ESTRO 2024 - Abstract Book

S4484

Physics - Machine learning models and clinical applications

ESTRO 2024

Figure 1: Overview of the system

We acquired MR images of C57BL/6J mice (n=25) that were given 66 Gy of fractionated radiotherapy covering the oral cavity, swallowing muscles and salivary glands over five days (two fractions per day) or twelve days (one fraction per day) [2]. Magnetic Resonance Imaging (MRI) of the H&N region was performed using a 7.05T BioSpec scanner (Bruker Medical systems, Germany) 3 days after irradiation. A fast T2 weighted spin-echo sequence, TurboRARE, withTE=31ms and TR=3100ms was employed. We included 420 and 450 MR images of irradiated and control animals, respectively, for further analysis. Data augmentation and image enhancement was done in terms of Contrast Limited Adaptive Histogram Equalization (CLAHE) [3] and notch frequency filtering [4]. All these images (original, CLAHE and notch filtered) were then assembled to form the dataset (see figure 1). After removing images of the most lateral parts of the animals (containing mostly low-intensity pixels of air), a balanced dataset with 900 images in each class, i.e. control and irradiated animals, was obtained. This dataset was then used as input to the multi head attention classifier [5].

Results:

Experiments were carried out using a NVIDIA RTX 2000 GPU with 12GB graphics memory. The dataset was divided into 75% training and 25% for testing. A pre-trained Vision Transformer (ViT) with twelve attention heads along with transfer learning [6] was used for our binary classification task. Given that ViT is a fairly large network (approximately 8.5 million parameters), a transfer learning strategy was adopted for reusing the most recent weights and adapting the last layer for the given task. We used a batch size of 32 with Adam optimizer. The classifier was trained for 40 epochs. The model was able to classify control and irradiated animals with an average accuracy of 81%. Table 1 depicts class wise confusion matrix and other parameters related to performance of the classifier.

Control

Irradiated

Control

170

55

Irradiated Precision

37

188

0.82 0.76 0.79

0.77 0.84 0.80

Recall

F1 score

Table 1: Confusion matrix (first 2 rows), precision, recall and F1 score class-wise.