ESTRO 2024 - Abstract Book

S4402

Physics - Intra-fraction motion management and real-time adaptive radiotherapy

ESTRO 2024

The overall workflow of the proposed TL-DSN-JC method is depicted in Figure 1 as the semantic segmentation of a tumor target on the kV projection 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 nets and combining them to form a jury committee (JC), we were able to determine whether a pixel belonged to the tumor target. Benchmarking against the ground truth, we defined three quantitative performance measures that were widely used to evaluate the classifier qualities: precision, recall and F1 scores. Among these three measures, F1, the harmonic mean of precision and recall, is a more balanced metric to evaluate the segmentation performance, which is also referred to as the Dice Similarity Coefficient.

Figure 1. Diagram indicating the pipeline of the proposed transfer learning deep segmentation net jury committee (TL-DSN-JC) method.

Results:

As shown in Figure 2, the precision (left), recall (middle), and F1 (right) measures for the proposed TL-DSN-JC method and the four other methods are provided. Specifically, the precision obtained with TL-DSN-JC outperformed the other four methods by a wide margin: on average it is around 0.84, compared to 0.78 of SAM (Segment Anything Model), 0.75 of TL-single-DSN, 0.62 of nTL-DSN-JC, and 0.44 of nTL-single-DSN, which indicates that the tumor regions declared by the proposed TL-DSN-JC are far more likely (84% on average) to be real tumor regions. Interestingly, in terms of recall, the proposed TL-DSN-JC (~0.88), was only better than nTL-DSN-JC (~0.85), and slightly worse than the two singular DSN methods: nTL-single-DSN (~0.94) and TL-single-DSN (~0.93), and SAM (~0.91) which means real tumor regions are enclosed by the single nets, but the single nets based method achieved this high percentage at the high cost of containing too many false positives shown in their considerably low precision values.

According to the F1 measure, the TL-DSN-JC had the best performance at ~0.85 as opposed to ~0.58 from the nTL single-DSN algorithm, an improvement of ~80%. The SAM indeed delivered impressive results: its median F1 measure