Paper deep dive
Missing Modality-Aware Calibration for Trustworthy Brain Tumor Segmentation
Sol Lee, Hyunji Kim, Sungrae Hong, Donghee Han, Mun Yi
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 94%
Last extracted: 10/10/2026, 1:32:38 AM
Summary
The paper introduces Missing Modality-Aware Local Temperature Scaling (MMA-LTS), a post-hoc calibration method for brain tumor segmentation models operating under incomplete multimodal MRI inputs. MMA-LTS addresses the limitation of existing calibration techniques that assume monotonic difficulty reduction with more modalities. Instead, it uses a spatially adaptive temperature field conditioned on a modality-availability learnable token and a voxel-wise difficulty score derived from Mahalanobis distance. Experiments on BraTS 2020 and FeTS 2024 demonstrate that MMA-LTS significantly improves Expected Calibration Error (ECE) while preserving segmentation accuracy (DSC) across various missing-modality scenarios.
Entities (12)
Relation Signals (10)
MMA-LTS → evaluatedon → BraTS 2020
confidence 98% · Experiments on BraTS 2020 and FeTS 2024 show that MMA-LTS improves calibration
MMA-LTS → evaluatedon → FeTS 2024
confidence 98% · Experiments on BraTS 2020 and FeTS 2024 show that MMA-LTS improves calibration
MMA-LTS → improves → Expected Calibration Error (ECE)
confidence 95% · Experiments on BraTS 2020 and FeTS 2024 show that MMA-LTS improves calibration... thereby enhancing trustworthiness
MMA-LTS → preserves → Dice Similarity Coefficient (DSC)
confidence 95% · MMA-LTS improves calibration while preserving the segmentation accuracy of state-of-the-art models
MMA-LTS → uses → Voxel-wise Difficulty Score (VDS)
confidence 92% · It estimates a spatially adaptive temperature field conditioned on a modality-availability learnable token and a voxel-wise difficulty score.
MMA-LTS → uses → Modality Availability Learnable Token
confidence 92% · It estimates a spatially adaptive temperature field conditioned on a modality-availability learnable token...
MMA-LTS → appliedto → DC-Seg
confidence 90% · We evaluate MMA-LTS on state-of-the-art brain tumor segmentation models... DC-Seg [10].
MMA-LTS → appliedto →
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Multimodal brain tumor segmentation typically leverages multiple MRI modalities, yet incomplete modality acquisition is common in clinical practice due to protocol heterogeneity and scan failures. Although recent methods maintain segmentation accuracy under missing modality conditions, they frequently overlook prediction reliability, leading to miscalibrated confidence estimates that hinder clinical adoption. Existing calibration techniques are largely modality-agnostic or assume that prediction difficulty decreases monotonically as additional modalities become available. However, in brain tumor segmentation, prediction difficulty depends primarily on which modalities are absent rather than how many, leading to combination-specific and spatially heterogeneous calibration errors. To address this, we propose Missing Modality-Aware Local Temperature Scaling (MMA-LTS), a post-hoc voxel-wise confidence calibration method. It estimates a spatially adaptive temperature field conditioned on a modality-availability learnable token and a voxel-wise difficulty score. Experiments on BraTS 2020 and FeTS 2024 show that MMA-LTS improves calibration while preserving the segmentation accuracy of state-of-the-art models across diverse missing-modality scenarios, thereby enhancing trustworthiness toward clinical deployment.
Tags
Links
- Source: https://arxiv.org/abs/2610.11419v1
- Canonical: https://arxiv.org/abs/2610.11419v1
Trouble viewing inline? Open PDF directly →
Full Text
43,378 characters extracted from source content.
Expand or collapse full text
Missing Modality-Aware Calibration for Trustworthy Brain Tumor Segmentation Sol Lee Hyunji Kim Sungrae Hong Donghee Han Mun Yi †thanks: Corresponding author. Affiliation: Korea Advanced Institute of Science and Technology, Daejeon, South Korea leesol4553, hyunjikim, sr5043, handonghee, munyi@kaist.ac.kr Abstract Multimodal brain tumor segmentation typically leverages multiple MRI modalities, yet incomplete modality acquisition is common in clinical practice due to protocol heterogeneity and scan failures. Although recent methods maintain segmentation accuracy under missing modality conditions, they frequently overlook prediction reliability, leading to miscalibrated confidence estimates that hinder clinical adoption. Existing calibration techniques are largely modality-agnostic or assume that prediction difficulty decreases monotonically as additional modalities become available. However, in brain tumor segmentation, prediction difficulty depends primarily on which modalities are absent rather than how many, leading to combination-specific and spatially heterogeneous calibration errors. To address this, we propose Missing Modality-Aware Local Temperature Scaling (MMA-LTS), a post-hoc voxel-wise confidence calibration method. It estimates a spatially adaptive temperature field conditioned on a modality-availability learnable token and a voxel-wise difficulty score. Experiments on BraTS 2020 and FeTS 2024 show that MMA-LTS improves calibration while preserving the segmentation accuracy of state-of-the-art models across diverse missing-modality scenarios, thereby enhancing trustworthiness toward clinical deployment. Keywords: Brain Tumor Segmentation Multi-Modal Missing Modality Confidence Calibration Post-hoc Calibration 1 Introduction Multimodal brain tumor segmentation plays a critical role in diagnosis, treatment planning, and prognosis assessment. Accurate delineation requires integrating complementary information from multiple MRI modalities, such as T1, T2, T1ce, and FLAIR [4]. However, in real-world clinical workflows, acquiring all modalities is often infeasible due to variations in imaging protocols, scan failures, or patient conditions [25]. As a result, models must operate robustly under incomplete modality settings. Prior approaches addressed this challenge by synthesizing missing modalities [16, 23] or training separate models for each possible modality combination [8, 22]. More recent methods achieve stronger performance by embedding multimodal inputs into a shared latent space [2, 10, 26], enabling a unified architecture. Despite recent advances, most studies prioritize segmentation accuracy while paying limited attention to prediction reliability under missing-modality conditions. Fig. 1(a) shows that the state-of-the-art DC-Seg [10] achieves higher Dice similarity coefficient (DSC) yet exhibits larger expected calibration error (ECE) than RobustSeg [2]. This suggests that high segmentation performance under incomplete multimodal inputs can obscure poorly calibrated confidence estimates that do not reflect true accuracy. Reliable confidence is essential in medical image segmentation, as miscalibration can undermine clinical decision-making [1]. This issue is particularly pronounced in brain tumor segmentation, where heterogeneous subregions and boundaries require calibrated confidence to identify unreliable delineations and support expert intervention [5, 15]. Figure 1: Segmentation performance and calibration results. (a) Left: DSC and ECE of state-of-the-art models on BraTS 2020, averaged over all modality combinations. Right: DC-Seg results per modality combination for WT. (b) Two cases with the same modality count but differing in combination, showing DC-Seg confidence maps (uncalibrated and MMA-LTS), voxel-wise difficulty score (VDS), and calibration strength (|T−1||T-1|). Most prior work on segmentation calibration relies on joint prediction-calib ration methods [1, 12, 18] or post-processing probability correction [5, 24], which assume fixed and fully observed inputs and therefore do not address multimodal or missing-modality settings. Although calibration methods for multimodal classification under missing modalities have been proposed [11, 14], they typically assume prediction difficulty decreases and confidence increases monotonically as more modalities become available. This assumption does not hold in brain tumor segmentation, where modalities contribute unequally: T1ce primarily highlights enhancing tumor (ET) regions, whereas T2 and FLAIR better capture whole tumor (WT) extent [17]. As illustrated in Fig. 1(b), prediction difficulty depends on which modalities are missing rather than how many, resulting in spatially heterogeneous calibration errors. Figure 2: The overall framework of MMA-LTS. Motivated by this observation, we propose MMA-LTS (Missing Modality-Aware Local Temperature Scaling), a plug-in post-hoc calibration method for multimodal segmentation with incomplete modalities. Unlike prior work, MMA-LTS performs voxel-wise calibration via modality-aware conditioning and spatial difficulty gating, correcting modality-combination-specific and spatially heterogeneous calibration errors without retraining. Experiments on BraTS 2020 [17] and FeTS 2024 [25] demonstrate improved calibration across missing-modality scenarios while preserving segmentation accuracy. 2 Method Assume disjoint data splits trainD_train, valD_val, and testD_test, each consisting of pairs (In,Yn)(I_n,Y_n), where n is the sample index, In∈ℝ|ℳ|×D×H×WI_n ^|M|× D× H× W is the multimodal input, Yn∈LD×H×WY_n∈ L^D× H× W is the ground-truth label, and ℳM and L denote the sets of modalities and classes, respectively. A segmentation model F pretrained on trainD_train takes InI_n and produces logits n∈ℝ|L|×D×H×Wz_n ^|L|× D× H× W. The proposed method aims to calibrate nz_n of F on testD_test under arbitrary modality availability. To this end, a calibration module is trained on valD_val, yielding spatially adaptive confidence estimates without retraining F. The overall framework is shown in Fig. 2. 2.1 Missing Modality-Aware Local Temperature Scaling Spatial Encoder. F extracts modality-specific features n,kk∈ℳ\f_n,k\_k and produces a joint representation ϕn∈ℝCf×D×H×W φ_n ^C_f× D× H× W and nz_n from InI_n, where CfC_f is the channel dimension. These two signals are concatenated and passed through a 3D CNN module EθE_θ to obtain a spatial embedding n∈ℝCf×D×H×Wh_n ^C_f× D× H× W, which serves as the base representation for temperature estimation. Voxel-wise Difficulty Score (VDS). To capture spatial difficulty induced by modality loss, we extend Mahalanobis-distance-based estimation [3] to 3D segmentation. We collect voxel feature-label pairs =(n,v,yn,v)V=\(g_n,v,y_n,v)\ from trainD_train to model the feature distribution of pretrained F, where v indexes a voxel location, n,v=ϕn(v)∈ℝCfg_n,v= φ_n(v) ^C_f denotes the feature vector at voxel v, yn,v=Yn(v)∈Ly_n,v=Y_n(v)∈ L the corresponding label, and lV_l the subset of voxels belonging to each class l∈Ll∈ L. We estimate the class-conditional and class-agnostic statistics as: l μ_l =1|l|∑(n,v)∈ln,v,=1||∑l∈L∑(n,v)∈l(n,v−l)(n,v−l)⊤, = 1|V_l| _(n,v) _lg_n,v, = 1|V| _l∈ L _(n,v) _l(g_n,v- μ_l)(g_n,v- μ_l) , (1) agn μ^agn =1||∑(n,v)∈n,v,agn=1||∑(n,v)∈(n,v−agn)(n,v−agn)⊤. = 1|V| _(n,v) g_n,v, ^agn= 1|V| _(n,v) (g_n,v- μ^agn)(g_n,v- μ^agn) . The class-conditional and agnostic Mahalanobis distances are then defined as: Mn,v=(n,v−yn,v)⊤−1(n,v−yn,v), M_n,v=(g_n,v- μ_y_n,v) ^-1(g_n,v- μ_y_n,v), (2) Mn,vagn=(n,v−agn)⊤(agn)−1(n,v−agn). M^agn_n,v=(g_n,v- μ_agn) ( ^agn)^-1(g_n,v- μ^agn). Mn,vM_n,v measures deviation from the class distribution and Mn,vagnM^agn_n,v from the global distribution, where yn,vy_n,v is obtained from ground-truth labels during training and replaced by the predicted label y^n,v=argmaxn,v y_n,v= _n,v at inference time. We adopt RMDn,v=Mn,v−Mn,vagnRMD_n,v=M_n,v-M^agn_n,v as a voxel-wise difficulty proxy, where larger values indicate voxels that are atypical within their class yet unremarkable globally, reflecting low class discriminability and thus indicating harder regions. After normalization and reshaping, we obtain VDSn∈[0,1]D×H×WVDS_n∈[0,1]^D× H× W. Modality Availability Learnable Token. We encode modality availability as a sample-level conditioning signal for calibration. Let (In)⊆ℳS(I_n) denote the set of available modalities and define the binary mask n,k=[k∈(In)]m_n,k=I[k (I_n)], where 1 and 0 indicate available and missing, respectively. For each k∈ℳk , learnable tokens kavail,kmiss∈ℝdee_k^avail,e_k^miss ^d_e represent the available and missing states, where ded_e is the token dimension. The modality state token n,kt_n,k is defined as n,k=n,kkavail+(1−n,k)kmiss.t_n,k=m_n,ke_k^avail+(1-m_n,k)e_k^miss. (3) Concatenating all modality tokens yields the summary vector ¯n=⨁k∈ℳn,k∈ℝ|ℳ|⋅de t_n= _k t_n,k ^|M|· d_e, which serves to globally modulate the calibration strength. Temperature Estimator. The estimator learns a mapping that produces a voxel-wise temperature field using spatial features nh_n, voxel-wise difficulty score VDSnVDS_n, and modality availability token ¯n t_n. Spatial difficulty gating is first applied to selectively emphasize regions requiring stronger calibration. Sample-level modality conditioning is then performed via FiLM-based affine modulation [21] to adjust feature modulation according to the sample-wise modality configuration. The resulting temperature field Tn∈ℝD×H×WT_n ^D× H× W is then given by Tn=Gθ((n⊙VDSn)⊙γn+βn),T_n=G_θ ((h_n _n) _n+ _n ), (4) where ⊙ denotes element-wise multiplication, (γn,βn)=fFiLM(¯n)( _n, _n)=f_FiLM( t_n) denote affine modulation parameters, and GθG_θ is a 3D CNN predictor. Spatial-difficulty gating determines where stronger calibration is needed, while sample-level modality conditioning controls how strongly it is applied based on available modalities. Local Temperature Scaling. Following prior work [5], the estimated temperature field TnT_n is applied to post-hoc rescale the logits nz_n of F via temperature-scaled softmax, yielding calibrated voxel-wise class probabilities: pcal,n,v(l)=σsoftmax(n,vTn,v)(l),l∈L.p_cal,n,v^(l)= _softmax ( z_n,vT_n,v )^(l),\;\;l∈ L. (5) where Tn,vT_n,v and n,vz_n,v denote the temperature and logit at voxel v, respectively. Since Tn,v∈ℝ+T_n,v ^+, class ordering is preserved, with values greater than 1 reducing overconfidence and values less than 1 mitigating underconfidence. Training Objective. ModDrop [20] is applied as ~n,k=δn,kn,k f_n,k= _n,kf_n,k, where δn,k _n,k is a Bernoulli indicator randomly set to 0 during training. MMA-LTS is optimized via the negative log-likelihood (NLL) loss over the voxel grid Ω , which counteracts the miscalibration of F by recovering the true probability distribution: ℒNLL=−1|Ω|∑v∈Ωlogpcal,n,v(yn,v)L_NLL=- 1| | _v∈ p_cal,n,v^(y_n,v) (6) 3 Experiments 3.1 Implementation Details Dataset. Experiments are conducted on BraTS 2020 [17] and FeTS 2024 [25], each providing 3D multi-contrast MRI scans with four modalities (FLAIR, T1ce, T1, T2) and three tumor subregions: whole tumor (WT), tumor core (TC), and enhancing tumor (ET). BraTS 2020 contains 369 cases split as (219/50/100) for (trainD_train/valD_val/testD_test) [10], and FeTS 2024 contains 1,251 cases split as (1000/125/ 126) [27]. All volumes are skull-stripped, co-registered to a common anatomical template, resampled to 1mm31m^3, and z-score normalized. Segmentation Backbones and Calibration Methods. We evaluate MMA-LTS on state-of-the-art brain tumor segmentation models for missing-modality robustness: RobustSeg [2], mmFormer [26], and DC-Seg [10]. For calibration, we compare against joint prediction-calibration approaches such as mL1-ACE [1] and SDC [12], as well as post-hoc probability correction methods, including Temperature Scaling (TS) [7], Dirichlet Scaling (DirS) [9], Local Temperature Scaling (LTS) [5], Selective Scaling (SelS) [24], and MC Dropout (MC) [6]. All calibration methods are applied uniformly across backbones. Training Settings. All backbones are trained on trainD_train with batch size 2 for 500/250 epochs on BraTS 2020/FeTS 2024, augmented with random 80380^3 crops, ±10∘± 10 rotation, intensity shifts, and flipping. Post-hoc calibration methods and MMA-LTS are trained for 40 epochs using AdamW [13] (batch size 11, learning rate 1×10−41× 10^-4) on valD_val, while mL1-ACE and SDC are trained with backbones. MC uses p=0.3p=0.3 with 5 forward passes. For MMA-LTS, we set Cf=16C_f=16 and de=256d_e=256. All experiments are conducted on an NVIDIA® RTX A6000 GPU. Evaluation Metrics. We evaluate segmentation with DSC [28] and calibration with ECE [19] following recent works [1, 12], where ECE is computed over foreground voxels using 10 bins and averaged across test volumes. Table 1: DSC and ECE comparison of DC-Seg on BraTS 2020. ∙ and ∘ denote available and missing modalities. Temperature scaling-based methods (TS, DirS, LTS, SelS, MMA-LTS) preserve the DSC of the uncalibrated model; only ECE is reported. Type FLAIR ∘ ∘ ∘ ∙ ∘ ∘ ∙ ∘ ∙ ∙ ∙ ∙ ∙ ∘ ∙ Avg. T1 ∘ ∘ ∙ ∘ ∘ ∙ ∙ ∙ ∘ ∘ ∙ ∙ ∘ ∙ ∙ T1ce ∘ ∙ ∘ ∘ ∙ ∙ ∘ ∘ ∘ ∙ ∙ ∘ ∙ ∙ ∙ T2 ∙ ∘ ∘ ∘ ∙ ∘ ∘ ∙ ∙ ∘ ∘ ∙ ∙ ∙ ∙ WT Uncalibrated 24.7224.72 21.7621.76 34.2034.20 26.6226.62 14.9414.94 19.6719.67 23.6523.65 24.0924.09 22.1422.14 13.6713.67 13.2813.28 21.4821.48 12.8312.83 15.0215.02 13.0013.00 20.0720.07 ⌞ DSC 85.0985.09 78.0978.09 76.1876.18 86.7986.79 87.2787.27 81.4181.41 89.4989.49 86.7486.74 89.5289.52 89.8689.86 90.6090.60 90.1490.14 90.4790.47 87.8187.81 90.8190.81 86.6886.68 mL1-ACE [1] 27.0427.04 25.4425.44 39.4139.41 26.4126.41 16.3716.37 23.4623.46 23.6923.69 25.9025.90 22.0822.08 13.9713.97 13.8013.80 21.2421.24 13.0013.00 16.4116.41 12.9812.98 21.4121.41 ⌞ DSC 83.7383.73 76.8576.85 72.4272.42 86.1286.12 86.5786.57 79.4179.41 88.2088.20 85.1385.13 88.4888.48 89.2289.22 89.9189.91 89.4789.47 89.9789.97 86.8386.83 90.4590.45 85.5285.52 SDC [12] 22.3522.35 21.7521.75 33.4233.42 23.4623.46 13.7213.72 19.1819.18 20.4220.42 21.8121.81 18.9918.99 12.6712.67 12.2212.22 18.3518.35 11.6811.68 13.8313.83 11.5811.58 18.3618.36 ⌞ DSC 85.6385.63 78.3478.34 75.5675.56 86.8186.81 87.7387.73 81.4681.46 89.5489.54 87.0587.05 89.7689.76 90.0790.07 90.7590.75 90.3890.38 90.7190.71 88.1188.11 91.0091.00 86.8686.86 MC [6] 24.7824.78 21.9521.95 34.1734.17 26.7226.72 14.9314.93 19.7419.74 23.5523.55 24.3024.30 22.1922.19 13.5713.57 13.3913.39 21.5221.52 12.9312.93 15.2515.25 12.9612.96 20.1320.13 ⌞ DSC 84.9684.96 77.9077.90 75.9675.96 86.7486.74 87.2887.28 81.2281.22 89.4189.41 86.4486.44 89.3089.30 89.8489.84 90.6090.60 90.2090.20 90.4290.42 87.6587.65 90.8990.89 86.5986.59 TS [7] 23.8523.85 21.0521.05 33.1633.16 25.7625.76 14.3314.33 18.9718.97 22.8422.84 23.2423.24 21.3521.35 13.1113.11 12.7412.74 20.7220.72 12.2812.28 14.4114.41 12.4612.46 19.3519.35 DirS [9] 24.8924.89 22.9022.90 34.7334.73 26.5726.57 15.6215.62 20.5920.59 23.3823.38 24.3224.32 22.0722.07 14.1114.11 13.5313.53 21.3521.35 13.2513.25 15.6115.61 13.2213.22 20.4120.41 LTS [5] 14.2314.23 16.2316.23 22.9122.91 16.2416.24 10.2810.28 14.2414.24 14.0014.00 13.9313.93 13.0713.07 10.1910.19 9.739.73 12.3512.35 9.479.47 10.3810.38 9.359.35 13.1113.11 SelS [24] 20.4320.43 18.1718.17 29.3329.33 22.9322.93 11.2711.27 15.2015.20 19.9119.91 19.8319.83 18.6518.65 10.6510.65 10.4010.40 18.0518.05 9.949.94 11.2311.23 10.1410.14 16.2916.29 MMA-LTS∗ 12.6412.64 12.7412.74 15.0815.08 12.9912.99 8.118.11 10.2210.22 11.7711.77 11.1811.18 11.4411.44 8.178.17 7.227.22 10.1210.12 6.956.95 7.37.3 6.776.77 10.1810.18 Uncalibrated 37.5637.56 19.0619.06 43.1643.16 43.3943.39 16.5316.53 17.9117.91 38.8938.89 35.6135.61 35.4935.49 16.4216.42 17.4017.40 35.3835.38 16.0816.08 16.9416.94 17.0117.01 27.1227.12 ⌞ DSC 73.2473.24 84.5284.52 64.6064.60 70.6770.67 87.5887.58 86.8386.83 73.9373.93 73.8573.85 75.3175.31 87.3687.36 87.4087.40 75.5175.51 87.4887.48 87.2087.20 87.3487.34 80.1980.19 mL1-ACE 33.6033.60 17.6417.64 40.7340.73 38.8338.83 13.9813.98 16.6316.63 34.2834.28 32.2332.23 31.0531.05 13.3513.35 14.2214.22 30.6530.65 12.9012.90 14.3114.31 13.7813.78 23.8823.88 ⌞ DSC 72.4572.45 84.0184.01 63.2263.22 70.0270.02 87.3587.35 84.8684.86 72.8172.81 73.0173.01 74.4374.43 87.5487.54 87.3887.38 74.8974.89 87.5587.55 86.7386.73 87.1387.13 79.5679.56 SDC 31.8131.81 16.7316.73 37.6337.63 36.2536.25 14.0314.03 15.0115.01 32.2032.20 30.1330.13 29.1129.11 13.9913.99 14.6814.68 29.0829.08 13.4813.48 14.4914.49 14.3614.36 22.8722.87 ⌞ DSC 73.2773.27 84.6584.65 64.3064.30 70.5170.51 87.5687.56 86.0486.04 74.1374.13 73.9973.99 75.8475.84 87.0687.06 87.1087.10 76.0176.01 87.2487.24 86.9286.92 87.0187.01 80.1180.11 TC MC 37.5637.56 19.3919.39 42.8442.84 43.6543.65 16.6416.64 17.8117.81 38.5638.56 35.7935.79 35.5735.57 16.3916.39 17.5017.50 35.2935.29 15.9815.98 17.0017.00 17.0417.04 27.1327.13 ⌞ DSC 72.9972.99 84.8384.83 54.8454.84 70.7070.70 87.5687.56 86.9386.93 74.3374.33 73.7073.70 75.3875.38 87.4787.47 87.5087.50 75.4575.45 87.4687.46 87.3587.35 87.3787.37 80.2680.26 TS 36.4736.47 18.4118.41 41.9141.91 42.2242.22 15.9315.93 17.2417.24 37.7937.79 34.5634.56 34.4434.44 15.7815.78 16.7616.76 34.3634.36 15.4715.47 16.3216.32 16.3916.39 26.2726.27 DirS 34.8634.86 17.6417.64 40.4840.48 40.6740.67 15.1615.16 16.3016.30 36.0536.05 32.9732.97 32.8532.85 15.0715.07 15.8115.81 32.7332.73 14.7514.75 15.4615.46 15.5215.52 25.0925.09 LTS 22.1622.16 12.5812.58 27.4827.48 27.7227.72 10.6010.60 11.0811.08 24.1024.10 21.1821.18 21.1721.17 10.1610.16 10.5210.52 21.5921.59 10.0410.04 10.9010.90 10.6710.67 16.8016.80 SelS 32.1632.16 15.6515.65 37.2037.20 39.7839.78 14.3214.32 15.4615.46 35.1135.11 31.9831.98 32.1532.15 13.9613.96 14.9814.98 31.9531.95 13.8513.85 14.6614.66 13.4813.48 24.1624.16 MMA-LTS∗ 18.1518.15 10.8610.86 20.3920.39 20.0720.07 10.210.2 10.0510.05 20.9420.94 18.8818.88 17.5817.58 9.689.68 10.3310.33 19.3419.34 9.79.7 10.6110.61 10.6210.62 14.4914.49 ET Uncalibrated 36.4536.45 12.8012.80 46.8446.84 46.8946.89 9.909.90 10.8810.88 43.3643.36 37.2237.22 37.4737.47 11.1611.16 10.4710.47 37.9637.96 10.4310.43 10.0610.06 10.3510.35 24.8224.82 ⌞ DSC 51.7651.76 79.9979.99 40.9240.92 42.4242.42 83.2483.24 83.3583.35 47.0247.02 51.6051.60 52.4552.45 82.6782.67 83.5183.51 52.5952.59 83.1783.17 83.5983.59 83.6483.64 66.8066.80 mL1-ACE 31.8331.83 11.0911.09 45.3145.31 42.9842.98 7.777.77 10.4510.45 39.5539.55 34.4034.40 32.8832.88 8.478.47 7.807.80 33.5833.58 7.907.90 7.967.96 7.937.93 21.9921.99 ⌞ DSC 51.3951.39 79.3579.35 38.5638.56 41.7841.78 83.2283.22 81.2181.21 45.8245.82 50.2550.25 51.8751.87 82.7482.74 83.5283.52 51.9751.97 83.1283.12 83.2783.27 83.3283.32 66.0966.09 SDC 11.8011.80 31.5231.52 42.7142.71 38.0238.02 8.858.85 9.709.70 36.1836.18 32.5932.59 31.3931.39 9.489.48 8.938.93 32.1232.12 8.858.85 8.888.88 9.009.00 21.3321.33 ⌞ DSC 51.3051.30 80.0080.00 39.0739.07 43.9743.97 83.3083.30 83.3283.32 47.6947.69 50.9650.96 54.6654.66 82.7782.77 83.4883.48 52.6952.69 83.2383.23 83.4883.48 83.4383.43 83.4883.48 MC 36.3736.37 13.0613.06 46.5846.58 47.5747.57 10.0910.09 10.5410.54 42.9942.99 37.4037.40 37.5637.56 11.1411.14 10.6110.61 37.6737.67 10.2910.29 10.0310.03 10.2810.28 24.8124.81 ⌞ DSC 51.4151.41 80.7580.75 40.5240.52 41.3641.36 83.3983.39 83.8483.84 46.6846.68 51.0651.06 51.8151.81 83.1683.16 83.6183.61 52.2852.28 83.6883.68 83.9583.95 83.7483.74 66.7566.75 TS 35.3035.30 12.3312.33 45.5445.54 45.6645.66 9.469.46 10.3810.38 42.2142.21 36.0836.08 36.3836.38 10.6710.67 9.999.99 36.8836.88 9.969.96 9.599.59 9.879.87 24.0224.02 DirS 33.2633.26 11.9611.96 43.4743.47 43.4743.47 9.089.08 10.0010.00 39.6939.69 33.9433.94 34.2234.22 10.2810.28 9.439.43 34.6334.63 9.539.53 9.099.09 9.309.30 22.7622.76 LTS 21.4521.45 9.369.36 30.6130.61 31.7031.70 7.477.47 7.607.60 29.2829.28 22.8922.89 23.4023.40 8.128.12 7.707.70 24.4224.42 7.747.74 7.587.58 7.717.71 16.4716.47 SelS 30.8730.87 11.2111.21 40.3840.38 42.7342.73 8.598.59 9.129.12 39.1539.15 32.9232.92 33.5133.51 9.769.76 9.209.20 33.9233.92 9.279.27 8.728.72 9.099.09 22.1022.10 MMA-LTS∗ 21.1421.14 8.848.84 25.125.1 26.9926.99 7.147.14 7.007.00 27.9727.97 22.8822.88 21.9321.93 7.787.78 6.856.85 23.3623.36 6.826.82 6.626.62 6.446.44 15.1215.12 Table 2: DSC and ECE comparison of DC-Seg on FeTS 2024. Type FLAIR ∘ ∘ ∘ ∙ ∘ ∘ ∙ ∘ ∙ ∙ ∙ ∙ ∙ ∘ ∙ Avg. T1 ∘ ∘ ∙ ∘ ∘ ∙ ∙ ∙ ∘ ∘ ∙ ∙ ∘ ∙ ∙ T1ce ∘ ∙ ∘ ∘ ∙ ∙ ∘ ∘ ∘ ∙ ∙ ∘ ∙ ∙ ∙ T2 ∙ ∘ ∘ ∘ ∙ ∘ ∘ ∙ ∙ ∘ ∘ ∙ ∙ ∙ ∙ WT Uncalibrated 20.8920.89 17.0017.00 25.2525.25 17.9217.92 12.0912.09 14.6614.66 15.7115.71 19.3019.30 15.3315.33 8.698.69 8.498.49 14.7314.73 8.498.49 12.0012.00 14.6014.60 15.0115.01 ⌞ DSC 87.6587.65 79.3379.33 80.8080.80 90.3190.31 88.9788.97 83.4483.44 91.8491.84 88.5388.53 91.9491.94 91.6491.64 92.2792.27 92.2292.22 92.5092.50 89.2089.20 88.8888.88 88.6388.63 mL1-ACE [1] 20.4420.44 20.7020.70 27.2027.20 15.9215.92 12.7012.70 17.4317.43 14.0014.00 18.5318.53 13.8413.84 8.128.12 8.128.12 13.1813.18 7.887.88 12.4612.46 7.967.96 14.5714.57 ⌞ DSC 86.4686.46 78.8278.82 78.4478.44 90.4690.46 88.0888.08 82.3482.34 91.7491.74 87.3487.34 91.8791.87 92.0892.08 92.4392.43 92.0592.05 92.6192.61 88.3588.35 92.6192.61 88.3888.38 SDC 18.1918.19 17.2617.26 22.6022.60 15.2315.23 10.8610.86 13.6113.61 13.1413.14 16.3016.30 13.1513.15 7.597.59 7.487.48 12.2512.25 7.347.34 10.7410.74 7.317.31 10.7410.74 ⌞ DSC 87.8487.84 80.1180.11 80.5080.50 90.7890.78 89.1689.16 83.6783.67 92.0992.09 88.6988.69 92.2592.25 92.1092.10 92.5792.57 92.4192.41 92.8492.84 89.3089.30 92.7892.78 89.1489.14 MC [6] 20.7620.76 16.9116.91 25.3225.32 17.9317.93 12.0312.03 14.6914.69 15.7615.76 19.2619.26 15.3315.33 8.688.68 8.528.52 14.7114.71 8.508.50 11.9811.98 14.5914.59 15.0015.00 ⌞ DSC 87.6887.68 79.2879.28 80.6380.63 90.3490.34 88.9988.99 83.3983.39 91.9791.97 88.4388.43 92.0792.07 91.7991.79 92.3792.37 92.3792.37 92.5692.56 89.0889.08 92.7092.70 88.9188.91 TS [7] 18.4318.43 15.3315.33 22.2122.21 15.5815.58 10.7710.77 13.1113.11 13.4813.48 16.9016.90 13.2213.22 7.597.59 7.407.40 12.6512.65 7.397.39 10.7310.73 12.8112.81 13.1713.17 DirS [9] 19.8519.85 18.0218.02 25.4825.48 16.2516.25 12.7012.70 15.5515.55 14.2814.28 18.4718.47 13.9913.99 8.498.49 8.478.47 13.5013.50 8.488.48 12.6012.60 14.3114.31 14.7014.70 LTS [5] 12.1912.19 14.0114.01 15.8915.89 12.0912.09 9.819.81 11.6811.68 10.0410.04 11.7711.77 9.909.90 8.788.78 8.648.64 9.699.69 8.758.75 10.0410.04 8.748.74 10.8610.86 SelS [24] 15.0715.07 13.5713.57 17.7917.79 13.0213.02 9.149.14 11.9511.95 11.1811.18 13.6813.68 11.0511.05 6.636.63 6.916.91 10.6810.68 7.027.02 8.988.98 10.8510.85 11.1711.17 MMA-LTS∗ 11.7811.78 13.2313.23 15.6515.65 11.8711.87 8.518.51 10.3010.30 9.899.89 10.2610.26 8.728.72 6.906.90 6.496.49 8.068.06 6.286.28 8.448.44 6.116.11 9.509.50 Uncalibrated 30.2330.23 14.4614.46 32.4032.40 31.8031.80 10.9010.90 12.9612.96 26.2626.26 25.9425.94 25.3525.35 11.6111.61 11.4811.48 23.6723.67 10.5910.59 10.8310.83 19.2719.27 19.8519.85 ⌞ DSC 75.7575.75 88.8988.89 73.8373.83 76.8676.86 91.5691.56 90.5090.50 80.3980.39 80.3680.36 81.0281.02 91.3391.33 91.8191.81 82.3082.30 92.0192.01 92.1592.15 85.4085.40 84.2984.29 mL1-ACE 26.4926.49 13.9413.94 29.2729.27 27.1827.18 9.879.87 12.1112.11 21.9821.98 22.9022.90 23.1923.19 10.6610.66 10.3110.31 21.3421.34 9.929.92 10.3610.36 9.709.70 17.2817.28 ⌞ DSC 74.2874.28 88.0488.04 71.4371.43 75.5375.53 90.8590.85 89.3689.36 78.7378.73 77.7077.70 78.6778.67 90.5690.56 90.9690.96 79.9679.96 91.1291.12 90.5690.56 91.2591.25 84.9384.93 SDC 23.4223.42 12.4912.49 28.6728.67 27.6327.63 9.669.66 11.1011.10 22.8722.87 21.9421.94 21.9021.90 10.0310.03 10.2110.21 20.2920.29 9.539.53 9.699.69 9.529.52 16.7316.73 ⌞ DSC 76.1876.18 89.1989.19 73.1373.13 77.0877.08 91.6091.60 90.5890.58 80.1980.19 80.0980.09 80.9380.93 91.6791.67 91.9091.90 82.3082.30 91.9191.91 92.0792.07 92.2292.22 85.4085.40 TC MC 30.1230.12 14.3514.35 32.7232.72 31.6631.66 10.8110.81 12.8712.87 26.2826.28 25.7025.70 25.1525.15 11.6611.66 11.4511.45 23.5723.57 10.6110.61 10.8610.86 19.2219.22 19.8019.80 ⌞ DSC 75.8975.89 88.9188.91 73.7973.79 76.9076.90 91.50091.500 90.4190.41 80.2980.29 80.4580.45 81.2281.22 91.3691.36 91.8591.85 82.6882.68 92.0092.00 92.1392.13 92.3892.38 85.4585.45 TS 26.8626.86 13.3113.31 28.5928.59 28.4128.41 9.809.80 11.8511.85 23.1323.13 22.7022.70 22.3822.38 10.5010.50 10.3810.38 20.8320.83 9.499.49 9.739.73 17.1617.16 17.6717.67 DirS 26.6726.67 13.2313.23 28.4828.48 28.2028.20 9.859.85 11.6511.65 23.0023.00 22.6822.68 22.1422.14 10.3410.34 10.2510.25 20.6920.69 9.399.39 9.599.59 17.0417.04 17.5517.55 LTS 18.1818.18 14.3814.38 19.8819.88 17.6917.69 11.6611.66 13.1813.18 15.3615.36 14.5014.50 14.7414.74 11.8711.87 11.7111.71 14.1614.16 11.2811.28 11.4211.42 11.3911.39 14.0914.09 SelS 23.6823.68 12.4712.47 23.2823.28 24.2724.27 9.229.22 11.5611.56 20.2720.27 19.5219.52 20.6720.67 9.839.83 9.799.79 18.9218.92 9.189.18 9.179.17 15.7615.76 15.8415.84 MMA-LTS∗ 17.7317.73 11.3011.30 19.5019.50 17.5617.56 7.837.83 9.359.35 15.1815.18 14.0514.05 14.2614.26 8.148.14 8.038.03 13.5613.56 7.537.53 7.557.55 7.787.78 11.9611.96 ET Uncalibrated 27.3527.35 10.1510.15 31.3831.38 27.8127.81 7.847.84 8.928.92 24.2024.20 23.6123.61 22.5622.56 7.677.67 7.517.51 21.4721.47 7.577.57 7.817.81 16.2216.22 16.8016.80 ⌞ DSC 58.8858.88 85.5185.51 54.6454.64 59.5759.57 87.6387.63 87.5587.55 63.5463.54 63.6763.67 65.2565.25 87.6587.65 88.6488.64 66.8266.82 88.0288.02 88.2988.29 75.6175.61 74.4874.48 mL1-ACE 23.8423.84 10.6610.66 28.7428.74 24.0224.02 7.517.51 8.908.90 20.3720.37 20.9620.96 21.0521.05 7.437.43 6.906.90 19.5219.52 6.926.92 7.687.68 6.536.53 14.7314.73 ⌞ DSC 57.9557.95 84.5684.56 52.5652.56 58.1458.14 87.0887.08 86.3486.34 62.3062.30 61.3761.37 63.8263.82 87.3987.39 88.1488.14 65.5165.51 87.8087.80 86.9386.93 88.2488.24 74.5474.54 SDC 23.0923.09 9.019.01 27.7627.76 23.8023.80 6.736.73 7.737.73 20.9520.95 19.7819.78 19.5219.52 6.876.87 6.806.80 18.4718.47 6.746.74 6.716.71 6.726.72 14.0514.05 ⌞ DSC 59.0959.09 85.4685.46 54.1354.13 59.2859.28 87.7487.74 87.5087.50 63.4463.44 63.4263.42 65.2065.20 87.7987.79 88.6088.60 66.8566.85 88.0988.09 88.3688.36 88.4688.46 75.5675.56 MC 27.5127.51 10.1210.12 31.6931.69 27.7227.72 7.747.74 8.808.80 24.3024.30 23.4823.48 22.4922.49 7.737.73 7.487.48 21.2921.29 7.647.64 7.817.81 16.2216.22 16.8016.80 ⌞ DSC 58.8658.86 85.5385.53 54.0054.00 59.2659.26 87.5387.53 87.5087.50 63.0463.04 63.5963.59 65.2665.26 87.7187.71 88.6188.61 66.8966.89 88.0388.03 88.2088.20 88.5488.54 75.5075.50 TS 24.1724.17 9.159.15 27.6227.62 24.6224.62 6.896.89 7.977.97 21.1521.15 20.5420.54 19.7519.75 6.756.75 6.576.57 18.8018.80 6.636.63 6.826.82 14.2614.26 14.7814.78 DirS 24.3324.33 9.329.32 27.8127.81 24.6724.67 7.147.14 7.987.98 21.3621.36 20.9020.90 19.7919.79 6.776.77 6.616.61 18.9218.92 6.706.70 6.896.89 14.3814.38 14.9014.90 LTS 20.2920.29 11.3511.35 21.7521.75 18.5618.56 9.869.86 9.639.63 17.8617.86 16.3416.34 16.4916.49 8.788.78 8.688.68 17.0317.03 8.588.58 8.568.56 8.348.34 13.4713.47 SelS 24.3724.37 8.788.78 24.8324.83 23.9623.96 6.456.45 7.687.68 20.0420.04 20.2320.23 18.9818.98 6.496.49 6.386.38 18.7118.71 6.186.18 6.466.46 13.8313.83 13.0513.05 MMA-LTS∗ 20.1020.10 8.628.62 21.6521.65 19.1419.14 5.925.92 6.686.68 17.8317.83 16.8416.84 15.6215.62 5.255.25 5.065.06 15.2615.26 5.035.03 5.505.50 5.085.08 11.5711.57 Table 3: DSC and ECE comparison of RobustSeg and mmFormer on BraTS 2020 and FeTS 2024. Values are reported as ECE/DSC and averaged over modality combinations. Dataset Model Type Uncal mL1-ACE SDC MC TS DirS LTS SelS MMA-LTS∗ BraTS 2020 RobustSeg WT 20.6820.68/84.0984.09 19.6819.68/80.1380.13 20.1020.10/83.5383.53 20.7320.73/84.3184.31 19.5419.54 20.4720.47 15.5715.57 15.9215.92 14.9314.93 TC 26.1826.18/75.3075.30 23.9623.96/71.3071.30 23.7323.73/74.8374.83 25.5025.50/75.4475.44 24.9024.90 23.9623.96 18.8918.89 18.6318.63 16.5616.56 ET 21.1821.18/62.9662.96 19.8719.87/59.7959.79 19.2819.28/62.7962.79 20.7620.76/63.1463.14 20.2220.22 20.1420.14 21.0221.02 20.0320.03 15.7815.78 mmFormer WT 22.1022.10/86.3286.32 20.3020.30/86.0686.06 20.3720.37/86.6286.62 22.1722.17/86.1386.13 21.4921.49 22.0522.05 17.0117.01 18.4018.40 11.6111.61 TC 29.1129.11/79.7079.70 25.2725.27/78.6578.65 26.3226.32/79.6579.65 29.1529.15/79.5679.56 28.4328.43 28.4328.43 19.3819.38 24.8824.88 17.4217.42 ET 26.5026.50/65.3565.35 22.5922.59/64.1564.15 23.9023.90/65.3065.30 26.6026.60/65.2065.20 25.8325.83 26.3126.31 17.7817.78 22.0522.05 16.4616.46 FeTS 2024 RobustSeg WT 15.9315.93/86.6686.66 17.6317.63/85.2885.28 15.0115.01/86.9186.91 16.5016.50/86.9186.91 13.9213.92 15.1915.19 11.4211.42 10.9010.90 9.459.45 TC 20.8420.84/81.6281.62 19.3619.36/79.4579.45 17.5817.58/81.6781.67 20.8820.88/81.6881.68 18.5018.50 18.5018.50 14.2614.26 14.8614.86 12.8912.89 ET 16.9716.97/71.7371.73 16.6316.63/69.7969.79 14.6914.69/71.1771.17 17.0917.09/71.7971.79 14.9314.93 15.2215.22 12.7712.77 13.1913.19 12.0312.03 mmFormer WT 16.6316.63/86.7586.75 14.5014.50/87.3687.36 11.8811.88/86.7486.74 16.8116.81/86.4686.46 15.3315.33 16.6616.66 12.7512.75 16.3416.34 11.8311.83 TC 22.8922.89/83.2583.25 15.5715.57/82.7982.79 15.4715.47/83.6483.64 22.9722.97/83.1483.14 21.4021.40 22.4822.48 16.8916.89 22.2922.29 13.2913.29 ET 19.5419.54/72.6972.69 13.7713.77/72.3272.32 13.9713.97/72.6372.63 19.6419.64/72.5872.58 28.0328.03 21.1721.17 17.2317.23 19.0219.02 13.1913.19 3.2 Experiment Results Comparative Experiments. Tables 1 and 2 present results on BraTS 2020 and FeTS 2024. MMA-LTS achieves the lowest ECE across most tumor regions and 15 modality combinations while preserving DSC. On BraTS 2020, MMA-LTS reduces average WT/TC/ET ECE to 10.2/14.5/15.1, outperforming the second-best LTS (13.1/16.8/16.5). On FeTS 2024, MMA-LTS similarly achieves the best ECE across tumor regions (9.5/12.0/11.6), demonstrating effective correction of modality-combination-induced and spatially heterogeneous confidence distortions. Table 3 further confirms consistent ECE reduction across RobustSeg and mmFormer backbones, validating its generalized calibration effectiveness. Ablation Study. Fig. 3 presents ablation results of MMA-LTS on DC-Seg with BraTS 2020. Removing VDS or replacing it with sample-averaged difficulty (SDSSDS) consistently increases ECE across all tumor regions, highlighting the importance of spatial difficulty estimation. Removing the modality availability token t¯n t_n or substituting it with a count-based token also degrades ECE, as it is limited in capturing modality combination-specific calibration errors. These results demonstrate that each component contributes to calibration improvement across all tumor regions under missing modalities. Figure 3: Ablation study of MMA-LTS. ECE are averaged over modality combinations. Figure 4: Time efficiency and sensitivity analysis (DC-Seg, BraTS 2020). (a) ECE versus inference and training time. (b) Sensitivity to CfC_f, ded_e, and valD_val size (dashed: 100%). Qualitative Analysis. Fig. 1(b) presents two cases with identical modality counts but different modality combinations (T2+FLAIR and T2+T1ce). Although their DSC values are comparable, the uncalibrated model exhibits different confidence patterns, with pronounced overconfident errors near tumor boundaries, assigning high tumor probability to non-tumor regions (false positives) and confidently mispredicting true tumors as background (false negatives), particularly in the T2+T1ce case. The VDS maps consistently highlight these ambiguous regions, while the calibration strength maps indicate where stronger correction is required for different modality combinations. MMA-LTS suppresses overconfident predictions and corrects spatially heterogeneous confidence distortions, yielding more reliable confidence estimates. Nonetheless, residual overconfidence near tumor boundaries remains, leaving room for improvement. Time Efficiency and Sensitivity Analysis. As a lightweight post-hoc module with a frozen backbone, MMA-LTS achieves the lowest ECE with inference time comparable to temperature scaling-based methods and much lower training cost than retraining approaches (mL1-ACE, SDC) (Fig. 4(a)). MC requires multiple forward passes and thus higher inference cost. Sensitivity analysis (Fig. 4(b)) confirms robustness to hyperparameters and valD_val size, showing stable ECE for Cf≥128C_f≥ 128 and de≥16d_e≥ 16, with performance saturating beyond 50% of valD_val. 4 Conclusion We present MMA-LTS, a plug-in post-hoc calibration method that corrects modality-combination-specific and spatially heterogeneous miscalibration in brain tumor segmentation under missing modalities. By conditioning on modality availability and voxel-wise difficulty, MMA-LTS enables spatially adaptive calibration without retraining the segmentation model or compromising accuracy. Experiments on BraTS 2020 and FeTS 2024 demonstrate consistent ECE reductions across multiple backbones and diverse missing-modality scenarios, establishing MMA-LTS as a practical step toward trustworthy clinical deployment. Future work includes extending validation to other anatomical domains and imaging modalities, incorporating reliability metrics beyond ECE with formal statistical testing, and further clinical validation in real-world settings. Acknowledgements This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korean government (MSIT) under grant number RS-2022-NR068758. We are also deeply grateful for the generous support provided by the Seegene Medical Foundation in South Korea. Disclosure of Interests. The authors have no competing interests to declare that are relevant to the content of this article. References [1] Barfoot, T., Garcia Peraza Herrera, L.C., Glocker, B., Vercauteren, T.: Average calibration error: A differentiable loss for improved reliability in image segmentation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. p. 139–149. Springer (2024) [2] Chen, C., Dou, Q., Jin, Y., Chen, H., Qin, J., Heng, P.A.: Robust multimodal brain tumor segmentation via feature disentanglement and gated fusion. In: International conference on medical image computing and computer-assisted intervention. p. 447–456. Springer (2019) [3] Cui, P., Zhang, D., Deng, Z., Dong, Y., Zhu, J.: Learning sample difficulty from pre-trained models for reliable prediction. Advances in Neural Information Processing Systems 36, 25390–25408 (2023) [4] Ding, Y., Yu, X., Yang, Y.: Rfnet: Region-aware fusion network for incomplete multi-modal brain tumor segmentation. In: Proceedings of the IEEE/CVF international conference on computer vision. p. 3975–3984 (2021) [5] Ding, Z., Han, X., Liu, P., Niethammer, M.: Local temperature scaling for probability calibration. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. p. 6889–6899 (2021) [6] Gal, Y., Ghahramani, Z.: Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In: international conference on machine learning. p. 1050–1059. PMLR (2016) [7] Guo, C., Pleiss, G., Sun, Y., Weinberger, K.Q.: On calibration of modern neural networks. In: International conference on machine learning. p. 1321–1330. PMLR (2017) [8] Hu, M., Maillard, M., Zhang, Y., Ciceri, T., La Barbera, G., Bloch, I., Gori, P.: Knowledge distillation from multi-modal to mono-modal segmentation networks. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. p. 772–781. Springer (2020) [9] Kull, M., Perello Nieto, M., Kängsepp, M., Silva Filho, T., Song, H., Flach, P.: Beyond temperature scaling: Obtaining well-calibrated multi-class probabilities with dirichlet calibration. Advances in neural information processing systems 32 (2019) [10] Li, H., Li, Z., Mao, Y., Ding, Z., Huang, Z.: Dc-seg: Disentangled contrastive learning for brain tumor segmentation with missing modalities. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. p. 138–148. Springer (2025) [11] Li, S., Chen, C., Han, J.: Simmlm: A simple framework for multi-modal learning with missing modality. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. p. 24068–24077 (2025) [12] Liang, W., Zhang, W.E., Yue, L., Xu, M., Maennel, O., Chen, W.: Calibrating on medical segmentation model through signed distance. In: Proceedings of the 34th ACM International Conference on Information and Knowledge Management. p. 1778–1787 (2025) [13] Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017) [14] Ma, H., Zhang, Q., Zhang, C., Wu, B., Fu, H., Zhou, J.T., Hu, Q.: Calibrating multimodal learning. In: International Conference on Machine Learning. p. 23429–23450. PMLR (2023) [15] Mehrtash, A., Wells, W.M., Tempany, C.M., Abolmaesumi, P., Kapur, T.: Confidence calibration and predictive uncertainty estimation for deep medical image segmentation. IEEE transactions on medical imaging 39(12), 3868–3878 (2020) [16] Meng, X., Sun, K., Xu, J., He, X., Shen, D.: Multi-modal modality-masked diffusion network for brain mri synthesis with random modality missing. IEEE Transactions on Medical Imaging 43(7), 2587–2598 (2024) [17] Menze, B.H., Jakab, A., Bauer, S., Kalpathy-Cramer, J., Farahani, K., Kirby, J., Burren, Y., Porz, N., Slotboom, J., Wiest, R., et al.: The multimodal brain tumor image segmentation benchmark (brats). IEEE transactions on medical imaging 34(10), 1993–2024 (2014) [18] Mukhoti, J., Kulharia, V., Sanyal, A., Golodetz, S., Torr, P., Dokania, P.: Calibrating deep neural networks using focal loss. Advances in neural information processing systems 33, 15288–15299 (2020) [19] Naeini, M.P., Cooper, G., Hauskrecht, M.: Obtaining well calibrated probabilities using bayesian binning. In: Proceedings of the AAAI conference on artificial intelligence. vol. 29 (2015) [20] Neverova, N., Wolf, C., Taylor, G., Nebout, F.: Moddrop: adaptive multi-modal gesture recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence 38(8), 1692–1706 (2015) [21] Perez, E., Strub, F., De Vries, H., Dumoulin, V., Courville, A.: Film: Visual reasoning with a general conditioning layer. In: Proceedings of the AAAI conference on artificial intelligence. vol. 32 (2018) [22] Qiu, Y., Chen, D., Yao, H., Xu, Y., Wang, Z.: Scratch each other’s back: Incomplete multi-modal brain tumor segmentation via category aware group self-support learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. p. 21317–21326 (2023) [23] Shen, L., Zhu, W., Wang, X., Xing, L., Pauly, J.M., Turkbey, B., Harmon, S.A., Sanford, T.H., Mehralivand, S., Choyke, P.L., et al.: Multi-domain image completion for random missing input data. IEEE transactions on medical imaging 40(4), 1113–1122 (2020) [24] Wang, D., Gong, B., Wang, L.: On calibrating semantic segmentation models: Analyses and an algorithm. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. p. 23652–23662 (2023) [25] Zhang, C., Chen, Y., Fan, Z., Huang, Y., Weng, W., Ge, R., Zeng, D., Wang, C.: Tc-diffrecon: Texture coordination mri reconstruction method based on diffusion model and modified mf-unet method. In: 2024 IEEE International Symposium on Biomedical Imaging (ISBI). p. 1–5. IEEE (2024) [26] Zhang, Y., He, N., Yang, J., Li, Y., Wei, D., Huang, Y., Zhang, Y., He, Z., Zheng, Y.: mmformer: Multimodal medical transformer for incomplete multimodal learning of brain tumor segmentation. In: International conference on medical image computing and computer-assisted intervention. p. 107–117. Springer (2022) [27] Zhu, S., Chen, Y., Chen, W., Wang, Y., Liu, C., Jiang, S., Qin, F., Wang, C.: Bridging the gap in missing modalities: Leveraging knowledge distillation and style matching for brain tumor segmentation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. p. 95–106. Springer (2025) [28] Zijdenbos, A.P., Dawant, B.M., Margolin, R.A., Palmer, A.C.: Morphometric analysis of white matter lesions in mr images: method and validation. IEEE transactions on medical imaging 13(4), 716–724 (1994)