A feasibility study of lung tumor segmentation on kilo-voltage radiographic images with transfer learning: Toward tumor motion tracking in radiotherapy.

PURPOSE: To segment the lung tumor on kilo-voltage X-ray radiographic images acquired during treatment toward the markerless lung tumor tracking.

METHODS

Per IRB approval, 1150 radiographic images from 80 lung cancer patients were included in the study. We developed a transfer learning deep segmentation net jury committee (TL-DSN-JC) algorithm to segment lung tumors on these images. The proposed models were initialized with the pre-trained VGG-16/19 networks with all but the weights of the connections between the final two layers frozen. A randomized partitioning was applied to train the deep segmentation net. By independently training 12 different deep segmentation nets (DSNs) to form a jury committee (JC), we could determine whether a pixel belonged to the tumor target. Meta-AI Segment Anything Model (SAM) was also tuned to cross-check with our proposed approach.

RESULTS

The results predicted by the TL-DSN-JC algorithm were evaluated using precision, recall, F1 score that is equivalent to the Dice score, and Hausdorff distance (HD). The TL-DSN-JC algorithm outperformed other similar algorithms such as the singular-DSN-without-transfer-learning, the DSN jury committee without transfer learning, and the singular-DSN-with-transfer-learning by up to 80%. Compared to the SAM-based model, the proposed model was superior in terms of HD although similar performance was observed based on the F1/Dice score. The results demonstrated that TL-DSN-JC could segment the tumor with clinically acceptable accuracy.

CONCLUSIONS

The experimental results demonstrated that the proposed algorithm outperformed the conventional deep learning techniques, offering a potential tool for markerless tumor motion tracking on projection images.