Performance of deep learning models for response evaluation on whole-body bone scans in prostate cancer

We aimed to develop deep learning classifiers for assessing therapeutic response on bone scans of patients with prostate cancer.

A set of 3791 consecutive bone scans coupled with their last previous scan (1528 patients) was evaluated. Bone scans were labeled as "progression" or "nonprogression" on the basis of clinical reports and image review. A 2D-convolutional neural network architecture was trained with three different preprocessing methods: 1) no preprocessing (Raw), 2) spatial normalization (SN), and 3) spatial and count normalization (SCN). Data were allocated into training, validation, and test sets in the ratio of 72:8:20, with the 20% independent test set rotating all scans over a five-fold testing procedure. A Grad-CAM algorithm was employed to generate class activation maps to visualize the lesions contributing to the decision. Diagnostic performance was compared using area under the receiver operating characteristics curves (AUCs).

The data consisted of 791 scans labeled as "progression" and 3000 scans labeled as "nonprogression." The AUCs of the classifiers were 0.632-0.710 on the Raw dataset, were significantly higher with the use of SN at 0.784-0.854 (p < 0.001 for Raw versus SN), and higher still with SCN at 0.954-0.979 (p < 0.001 for SN versus SCN). Class activation maps of the SCN model visualized lesions contributing to the model's decision of progression.

With preprocessing of spatial and count normalization, our deep learning model achieved excellent performance in classifying the therapeutic response of bone scans in patients with prostate cancer.