Order-Aware 2.5D Multiple Instance Learning for Preoperative MRI-Based Perineural Invasion Risk Assessment in Intrahepatic Cholangiocarcinoma

I. INTRODUCTION

• PNI is clinically relevant in ICC because surgical studies associate it with recurrence and worse survival, yet it is usually established by postoperative pathology.Routine MRI lacks nerve-level or slice-level labels for model training.Source: §I · p. 1
• Multiple instance learning fits this setting by supervising a bag of image instances with one patient-level label.Attention-based MIL provides differentiable bag-level pooling without local annotations.Source: §I · p. 1
• Ordered 2.5D slabs retain local cross-slice information while avoiding the higher memory and optimization burden of fully volumetric encoders.Unlike exchangeable instances, axial slab order can encode anatomical continuity.Source: §I · p. 1
• OAS-MIL models preoperative PNI prediction from tumor-centered T2-weighted MRI under patient-level supervision.The framework represents each crop as a complete ordered sequence of contiguous slabs.Source: §I · p. 1
• The sequence-attention branch uses a bidirectional recurrent reader to model axial order alongside the existing set-attention branch.The evaluation uses patient-level, five-fold label-stratified cross-validation with ablations of representation and aggregation choices.Source: §I · p. 1

• Each patient is represented by a tumor-centered T2-weighted MRI crop X_i with a binary patient-level PNI label y_i.Source: A. Problem Formulation and Preprocessing · p. 2
• Crops are resampled to 1.0×1.0×5.0 mm spacing and standardized to a 192 × 192 × 160 mm field of view.This yields H = W = 192 and D = 32.Source: A. Problem Formulation and Preprocessing · p. 2
• Each axial position forms an odd-sized slab from contiguous slices, with boundary indices clamped to preserve one slab per position.Source: B. Ordered 2.5D Slab Construction · p. 2
• The resulting patient bag retains the original through-plane order, with k = 3 used unless otherwise specified.Source: B. Ordered 2.5D Slab Construction · p. 2

• A shared adapted ResNet-18 backbone encodes every slab position into a feature vector of dimension d = 512.The same encoder is applied across all slab positions.Source: C. Shared Slab Encoder · p. 2
• The pretrained ResNet-18 first convolution is modified to accept the slab’s channel size through grayscale-based symmetric initialization.Source: C. Shared Slab Encoder · p. 2
• This initialization preserves symmetry across slab channels while keeping first-layer response scale comparable across slab sizes.Source: C. Shared Slab Encoder · p. 2

• The branch applies additive attention pooling over the slab bag and normalizes scalar scores across instances.Source: D. Set-Attention Branch · p. 2
• Its output is a set-attention summary that aggregates the slab representations.Source: D. Set-Attention Branch · p. 2
• Because the branch is permutation-invariant, its pooled representation does not depend on slab ordering.Source: D. Set-Attention Branch · p. 2

E. Sequence-Attention Branch

• The branch models slab features as an axial sequence using a one-layer bidirectional GRU to produce contextual states.Source: E. Sequence-Attention Branch · p. 2
• Bidirectional states are concatenated and projected to the common feature dimension before attention scoring.Source: E. Sequence-Attention Branch · p. 2, Figure 2 · p. 3
• The branch computes aggregation weights and a weighted summary from the projected sequence states.Source: Figure 2 · p. 3, E. Sequence-Attention Branch · p. 3
• Attention scores and normalized through-plane weights are computed from the projected sequence states.Source: E. Sequence-Attention Branch · p. 3, Figure 2 · p. 3

F. Prediction Head and Training Objective

• Branch summaries are normalized with shared Layer-Norm, concatenated, and passed through a prediction head.Source: F. Prediction Head and Training… · p. 3
• Training uses positive-class-weighted binary cross-entropy with logits, with the positive-class weight computed from the training split.Source: F. Prediction Head and Training… · p. 3
• Dropout is applied during training in the feed-forward components.Source: F. Prediction Head and Training… · p. 3

A. Data and Evaluation Protocol

• The cohort included 183 patients with preoperative T2-weighted MRI, comprising 70 PNI-positive and 113 PNI-negative cases.Source: A. Data and Evaluation Protocol · p. 3
• Five-fold label-stratified cross-validation was performed at the patient level, with separate training, checkpoint-selection, and held-out testing folds.Source: A. Data and Evaluation Protocol · p. 3
• Volumetric baselines used preprocessed crops, while MIL baselines used the same ordered 2.5D slab instances.Source: A. Data and Evaluation Protocol · p. 3
• AUROC was the primary metric and AUPRC the complementary metric for the imbalanced endpoint.Source: A. Data and Evaluation Protocol · p. 3
• OAS-MIL achieved AUROC 0.770 ± 0.077 and AUPRC 0.692 ± 0.093, versus CLAM-SB AUROC 0.705 ± 0.071 and AUPRC 0.608±0.118.Source: C. Comparison with Baselines · p. 3, Table I · p. 3

D. Ablation Analysis

• The 3-slice slab representation, fused set- and sequence-attention aggregation, and bidirectional GRU reader produced the strongest overall ablation results.Source: D. Ablation Analysis · p. 3, Table II · p. 4
• With test-time crop perturbations, OAS-MIL achieved AUROC 0.759 ± 0.069 and AUPRC 0.682±0.052, remaining close to unperturbed performance.Source: E. Crop Localization Sensitivity · p. 4
• The perturbation analysis independently varied in-plane offsets from U(−10, 10) mm and through-plane offsets of −1, 0, or +1 axial slices without retraining.Source: E. Crop Localization Sensitivity · p. 4
• The computational profile reports model size, multiply–accumulate operations, peak GPU memory, and forward time.Source: F. Computational Profile · p. 4, Table III · p. 4

• Aggregation weights concentrate over a limited slab interval near the tumor-containing region in the PNI-positive example.Source: G. Qualitative Analysis · p. 4
• The bottom panel pairs the selected slab with a tumor-mask reference overlay and a Grad-CAM overlay showing a coarse response around the tumor and adjacent tissue.Source: G. Qualitative Analysis · p. 4, Figure 3 · p. 4
• The top panel displays branch-wise aggregation-weight profiles across the ordered 2.5D slab montage.Source: Figure 3 · p. 4
• These outputs serve as qualitative audit cues rather than voxel-level PNI localization.Source: G. Qualitative Analysis · p. 4