STL-DeepBDC: A Robust Few-Shot Learning Framework for Multiclass Ovarian Tumor Classification in Ultrasound Outperforms Conventional Transfer Learning.
Authors
Affiliations (3)
Affiliations (3)
- Department of Ultrasound, The Fourth Affiliated Hospital of Harbin Medical University, Harbin, Heilongjiang, China.
- School of Computer Science and Technology, Harbin Institute of Technology, Harbin, China.
- In-Patient Ultrasound Department, The Second Affiliated Hospital of Harbin Medical University, Harbin, China.
Abstract
Accurate preoperative pathological classification of malignant and borderline ovarian tumors (OMTs) can support surgical planning, fertility preservation, and prognosis, but ultrasound-based subtype assessment remains difficult because imaging phenotypes are heterogeneous, class imbalance is common, and interpretation varies among operators. This study developed and evaluated an exploratory deep learning framework for multiclass ovarian tumor classification under small-sample conditions. This single-center retrospective study used ultrasound images from patients with surgically and pathologically confirmed OMTs treated between August 2017 and October 2022. Six pathological types were included. After quality control, preprocessing, and augmentation, the dataset was split at the patient level into training, validation, and test sets at a 5:3:2 ratio. An optimized ConvNeXt model served as the baseline. STL-DeepBDC combined stratified transfer learning with a Deep Bilinear Discriminative Covariance module. Performance was evaluated using accuracy, sensitivity, specificity, macro-F1 score, AUC, AUPRC, and Grad-CAM. STL-DeepBDC outperformed the optimized ConvNeXt baseline. Accuracy increased from 0.51 to 0.67, macro-average sensitivity from 0.53 to 0.68, specificity from 0.87 to 0.92, and macro-F1 score from 0.52 to 0.66. Macro-average AUC improved from 0.75 to 0.87, and macro-average AUPRC from 0.49 to 0.63. Category-level AUCs ranged from 0.76 to 0.94 and AUPRCs from 0.51 to 0.78, suggesting more stable discrimination across six pathological types under class imbalance. Ablation analysis supported the contributions of STL and DeepBDC. Grad-CAM highlighted clinically relevant regions, including thickened cyst walls, papillary projections, septations, and solid components. STL-DeepBDC provides an exploratory, interpretable framework for ultrasound-based multiclass ovarian tumor classification in a limited-sample setting. Larger multicenter studies with external validation are required before clinical implementation.