CutMix-augmented subtype classification in gastric spindle cell tumour

초록

Background & objectives: The tumour subtypes are crucial for the decision of treatment strategy in clinical practice. For deep learning-based approaches, there are challenges with an imbalanced distribution across subtypes. Here, we propose a simple and effective augmentation strategy to mitigate this. Methods: In this study, we retrospectively reviewed and collected the whole slide images of gastric spindle cell tumours from three hospitals: gastrointestinal stromal tumour (n = 502), leiomyoma (n = 145), and schwannoma (n = 35). We performed CutMix augmentation on classes with lower ratios, creating additional patches. Further, Uniform Manifold Approximation and Projection feature visualization was applied. Results: Through the augmentation strategy, the trained model achieved a betterIt is expected that our pipeline would offer significant support in clinical environments for the categorization of different subtypes of gastric cancer F1-score from 0.977 to 0.989 for GIST, from 0.720 to 0.744 for leiomyoma, and from 0.763 to 0.791 for schwannoma. Conclusion: Our study confirmed the effectiveness of the deep learning approach in challenging tasks involving a subtype classification of gastric spindle cell tumours and validated the efficacy of a simple augmentation strategy.