ESTRO 2024 - Abstract Book

S5331

Radiobiology - Tumour biology

ESTRO 2024

high nuclei concentration using a conventional stain separation method. Blurry patches were removed. This resulted in 2207 samples for exploration and 639 for the test set.

Modeling: We considered a binary classification problem using the two categories radioresistant and radiosensitive. DenseNet121, a CNN that was originally trained on real-world imaging data, was adapted to our classification task. We replaced its fully connected layers with 3 new ones consisting of 512, 128 and 1 neurons, respectively, with a sigmoid activation function at the output. Two experiments were conducted: (A) 5-fold cross validation on the exploratory cohort, ensuring a balanced split with respect to the tumor models, and subsequent validation on the test set. (B) five training iterations on the exploratory cohort, keeping one resistant and one sensitive model out, and subsequent validation on these two models, to further explore the CNN’s generalization capability. Predictive performance was assessed at the patch and the slide level, by computing accuracy, precision and recall. Slide-level predictions were obtained through majority voting, where each slide was assigned the most frequently predicted label among its patches as shown in Figure 1.

Results:

In experiment A, the CNN achieved average patch-wise metrics of 0.94 for accuracy, 0.92 for precision and 0.93 for recall for the validation folds on the exploration set, while all slide-wise metrics achieved a perfect score of 1. The CNN maintained a high performance on the test set with average patch-wise metrics of 0.87, 0.91 and 0.82, and average slide-wise metrics of 0.93, 1.00 and 0.87 for accuracy, precision and recall, respectively. In experiment B, the CNN’s performance decreased to a still acceptable level, with averaged patch-wise metrics of 0.79, 0.77 and 0.83, and averaged slide-wise metrics of 0.85, 0.84 and 0.90 for accuracy, precision and recall, respectively.