Contour-free multi-instance learning versus conventional CT features for preoperative Ki-67 stratification in gastric gastrointestinal stromal tumors: a retrospective comparison

Technical Summary: Contour-free Multi-instance Learning versus Conventional CT Features for Preoperative Ki-67 Stratification in Gastric GISTs

Problem Statement
Gastrointestinal stromal tumors (GISTs) exhibit significant biological variability, and preoperative risk assessment is often limited because key prognostic factors, such as the mitotic count and tumor rupture status, typically require intraoperative or pathologic confirmation. The Ki-67 labeling index serves as a critical measure of tumor proliferation, with higher values correlating to aggressive disease and recurrence. However, Ki-67 assessment relies on biopsy or resected tissue and is subject to sampling heterogeneity. While contrast-enhanced computed tomography (CT) is routinely used for staging, existing predictive models often rely on manually delineated tumor regions of interest (ROIs). This manual contouring introduces a separate preprocessing step, increases workload, and limits the scalability of radiomic or deep learning approaches. Furthermore, a single patient-level Ki-67 label corresponds to a full CT series containing hundreds of slices, only a subset of which depict the tumor. This "one label, many instances" structure is not optimally addressed by standard supervised learning without segmentation.

Methodology
This retrospective single-center study evaluated a contour-free Multi-Instance Learning (MIL) framework designed to stratify gastric GISTs by Ki-67 index without requiring tumor segmentation.

• Cohort and Tasks: The study included 61 patients with gastric GISTs who underwent radical resection. Three binary classification tasks were defined based on Ki-67 cutoffs of 5%, 8%, and 10%. Patients with interval-reported Ki-67 values spanning a specific cutoff were excluded only from that specific task.
• Data Processing:
  • Conventional Features: Two readers independently extracted 21 conventional CT variables (demographics, tumor morphology, enhancement patterns, necrosis, etc.) from arterial and venous phase images.
  • Contour-free MIL: Complete arterial and venous CT series were processed without tumor localization. DICOM slices were ordered by physical position, windowed to an abdominal setting, and standardized. A frozen ResNet-18 encoder (pretrained on ImageNet) extracted 512-dimensional feature vectors from each slice.
  • MIL Architecture: Each patient's full CT series formed a "bag" of instances (slices). Five MIL methods were implemented to aggregate slice features into a patient-level representation: MeanMIL, Attention-based MIL (ABMIL), Gated-attention MIL (GAMIL), Dual-stream MIL (DSMIL), and Attention-stabilized MIL (ASMIL).
• Model Comparison: The study compared:
  1. Two conventional models (Logistic Regression with L2 and Elastic Net regularization).
  2. Five MIL methods applied separately to arterial and venous phases.
  3. Four equal-logit fusion combinations (e.g., Arterial + Venous, Conventional + Arterial, etc.).
• Evaluation Protocol: A patient-level stratified 5-fold cross-validation, repeated 10 times, was used. Performance was assessed via AUROC, AUPRC, calibration, and Decision Curve Analysis (DCA). Statistical significance of differences was determined using paired bootstrap tests with Holm correction. A post-training "attention audit" was conducted to verify if learned attention weights correlated with tumor-bearing slices using manually annotated ROIs (which were not used during training).

Key Results

• Discriminative Performance: The study found no clear discrimination advantage for contour-free MIL or fusion models over conventional CT features.
  • The highest AUROCs were 0.645 (5% cutoff), 0.646 (8% cutoff), and 0.605 (10% cutoff).
  • Crucially, every AUROC confidence interval included 0.5, indicating that none of the models demonstrated statistically significant discrimination above chance in this cohort.
  • After Holm correction, none of the 90 pairwise comparisons between models, nor the 54 comparisons between fusion models and their components, remained statistically significant.
• MIL Method Comparison: Among the five MIL methods, only two specific pairwise comparisons remained significant after correction (Venous DSMIL vs. ASMIL at 8% AUROC; Venous DSMIL vs. GAMIL at 10% AUPRC). No MIL method consistently outperformed the others or the conventional models.
• Attention Audit: In the post-training audit, 34 of 60 tests showed significant association between attention weights and tumor-bearing slices after global Holm correction. However, the direction of this association varied (21 above reference, 13 below), suggesting that attention weights optimized for patient-level classification do not consistently localize the tumor.
• Stability: Fusion components showed low to moderate correlation, and the three-component fusion was less variable than single-phase MIL models but more variable than conventional elastic net.

Significance and Claims
The paper claims to define a patient-level framework for evaluating complete CT series without lesion delineation, offering a potential alternative to the labor-intensive manual contouring required by traditional radiomics.

However, the authors are explicit about the limitations of their findings:

1. No Clinical Superiority: In this specific cohort, contour-free MIL did not outperform conventional CT features. The authors state that the small sample size and wide confidence intervals preclude a firm conclusion that the approaches have equivalent performance, but they certainly did not demonstrate superiority.
2. Annotation Burden: The study did not measure the reduction in annotation time or clinical workload, which is a primary motivation for contour-free approaches.
3. Future Requirements: The authors conclude that external validation and direct assessment of annotation burden are required before this design can be considered for clinical use. They suggest future studies should utilize larger cohorts, CT-specific representations, and spatial registration between arterial and venous phases.

In summary, while the study successfully implemented a contour-free MIL pipeline for Ki-67 stratification, it failed to demonstrate that this approach provides a discrimination advantage over simpler, conventional CT feature analysis in the context of gastric GISTs.