Figure 2.

Schematic of two deep learning frameworks. The primary tumor model uses the primary tumor in CE-T1WI and T2WI sequences as the VOI with manual segmentation by radiologists. Contrastingly, the whole abdomen model uses the entire abdomen volume as VOI without any segmentation or delineation by hand. VOI on CE-T1WI and T2WI sequences were prepared for (A) pre-processing, which consisted of image segmentation (or not), registration, and normalization; (B) the backbone of the pre-trained 3D ResNet network was transferred to extract features and a global average pooling layer was added so that 1,024 features could be extracted from each patient; (C) the PCA was used for decomposition; (D) the platinum sensitivity prediction model was constructed using SVM; (E) the feature map that the last convolutional layer output was generated as heat maps for visualization. CE-T1WI, contrast-enhanced T1-weighted imaging; T2WI, T2-weighted imaging; VOI, volume of interest; MRI, magnetic resonance imaging; PCA, principal component analysis; SVM, support vector machine.