Research indexObservational
Weakly Supervised Multiple Instance Learning for Benign–Malignant Classification of Parotid Gland Tumors on Whole-Slide Images
Lee, Rao, Li and 4 others
21 Oct 2026Open access
NobleIDark:/48914/w1/20261119/C62D6701DOI10.65215/ltspreprints.2026.10.20.000365
Weakly Supervised Multiple Instance Learning for Benign–Malignant Classification of Parotid Gland Tumors on Whole-Slide Images
Lee, Rao, Li and 4 others · 2026
NobleIDark:/48914/w1/20261119/C62D6701DOI10.65215/ltspreprints.2026.10.20.000365
Abstract
Background Accurate differentiation between benign and malignant parotid gland tumors (PGTs) is important for diagnosis and treatment planning. Although histopathologic examination of hematoxylin and eosin (H&E)-stained slides remains the diagnostic reference standard, whole-slide image (WSI) assessment is labor-intensive and may be challenging in histologically heterogeneous lesions. Weakly supervised multiple instance learning (MIL) enables WSI-level classification using case-level diagnostic labels without requiring region-level annotation. Methods We retrospectively analyzed 101 patients with PGTs, including 58 benign and 43 malignant cases, represented by 189 digitized WSIs. Patient-level stratified five-fold cross-validation was performed to ensure that all WSIs from an individual patient were assigned exclusively to one fold. Following automated tissue detection and quality-controlled patch extraction, 256 × 256-pixel patches were encoded using either ResNet-50 or the pathology-pretrained CTransPath backbone. Three MIL strategies—CLAM-SB, attention-based MIL (ABMIL), and mean pooling—were evaluated. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, F1-score, and area under the precision–recall curve. Results CTransPath + CLAM‑SB yielded the highest mean AUC among the evaluated configurations (0.83), with an accuracy of 0.81 sensitivity of 0.82, specificity of 0.80, and F1‑score of 0.80. Under the CLAM‑SB framework, CTransPath showed a higher mean AUC than ResNet‑50 (0.83 vs. 0.81). Using CTransPath features, CLAM‑SB achieved higher mean AUC than ABMIL (0.80) and mean pooling (0.77). However, pairwise differences between models did not reach statistical significance. The optimized CLAM‑SB configuration with a bag loss weight of 0.3 further increased the mean AUC to 0.86. Conclusions Weakly supervised MIL enabled benign–malignant classification of PGTs from routine H&E WSIs without region-level annotation. In this retrospective single-center cohort, CTransPath combined with CLAM-SB showed the most favorable overall performance, with a consistent trend toward better discrimination across models. Given the limited sample size, these findings should be regarded as preliminary and warrant validation in larger, multicenter cohorts with external validation to confirm their generalizability and clinical utility.