Abstract
Following uveal melanoma (UM) affected treatment using ruthenium-106 brachytherapy, tumor thickness patterns fall into one of four categories: decrease (regression), increase (recurrence), stop (stable), or other, which are assessed in follow-up A-mode and B-mode images. These patterns are critical indicators of the tumor's response to therapy. This study aims to apply deep learning (DL) models for predicting post-brachytherapy tumor response patterns. A cohort of 192 patients participated in this study. B-Mode images taken at the time of diagnosis were collected, and the ophthalmologists labeled the images into four response patterns based on the results of the treatment. DenseNet121 and ResNet34 models were trained and evaluated using performance metrics. DenseNet121 achieved a macro-average AUC of 0.933 (0.95% CI [0.905-0.957]), compared to 0.916 (95%CI [0.884-0.945]) for the ResNet34. The per-class evaluation showed that DenseNet121 excelled in predicting all categories, providing superior predictive accuracy. This difference in classification performance was statistically significant based on the DeLong test (p < 0.05). The ablation study revealed that the best performance was achieved without pretrained weights, using dropout layers and a batch size of 32. Both models demonstrated strong classification capabilities, with DenseNet121 providing the highest overall accuracy. This study highlights the potential of DL models in predicting response patterns in UM patients undergoing brachytherapy. Further validation and exploration of their integration into clinical practice are warranted.