Abstract 5427: Improving non-small cell lung cancer segmentation on a challenging dataset

Abstract

When applied to different datasets, performance of the same deep learning tumor segmentation model can greatly vary. In a non-small cell lung cancer CT scan segmentation study that consists of two datasets, we found that the SwinUNETR model achieves state-of-the-art DICE score on a public dataset NSCLC but performs badly on a private dataset of curated data collected clinically. This performance variation reduces the applicability of such models. To mitigate this gap, through experimentation, we identified a set of techniques and applied them in the following order: (1) normalize a dataset to reduce differences between images. (2) stratify a dataset according to tumor sizes to form a more diverse training set. (3) isolate the lung area before training to help the model focus on the right area. (4) before training, initialize models with self-supervised pre-training weights (5) use a new loss function to give more weights on the cancerous area (6) after a model is trained, perform 3-axis test time flipping augmentation and ensemble the final predictions. In our experiments, our set of techniques improved the test DICE score for both datasets we tested on, where the best improvement was a 53% improvement from 0.32 to 0.49 DICE score. Citation Format: Yi Wei, Balaji Selvaraj, Mayank Patwari, Qin Li, Meng Xu, Konstantinos Sidiropoulos, Zhenning Zhang, Leon Fedden, Anant Madabhushi, Mohammadhadi Khorrami, Vidya Sankar Viswanathan, Amit Gupta. Improving non-small cell lung cancer segmentation on a challenging dataset. [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2023; Part 1 (Regular and Invited Abstracts); 2023 Apr 14-19; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2023;83(7_Suppl):Abstract nr 5427.