ESTRO 2024 - Abstract Book

S4967

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2024

for classification. VGG-16, GoogLeNet, UNet, AlexNet, DenseNet, and ResNet34 are selected for evaluation due to their proven effectiveness in image classification tasks. To build DL models, we used Python 3.9.12, with TensorFlow 2.9.1, and Keras 2.9.0 framework for neural networks. Classification performance is assessed using accuracy, F-score, sensitivity, specificity, and area under the receiver operating characteristic curve (AUC). A confusion matrix will be constructed to visualize classification results. The dataset is divided into training (75%), validation (10%), and testing (15%) subsets. Comprehensive hyperparameter tuning was conducted to optimize model performance based on response surface methodology (RSM).

Results:

In our study, we investigated the classification performance of various 3D neural network architectures for a specific medical task. The accuracy results revealed that GoogleNet achieved an accuracy of 74%, while DenseNet reached 80%, and ResNet34 performed even better with an accuracy of 92%. Unfortunately, other network architectures did not achieve acceptable levels of accuracy on our dataset. Notably, our state-of-the-art network outperformed all others, achieving an impressive accuracy of 94%, establishing it as the leading model in terms of accuracy. We also evaluated the models using the area under the receiver operating characteristic curve (AUC). GoogleNet exhibited an AUC of 0.74 (95% CI, 0.690–0.792), DenseNet had an AUC of 0.84 (95% CI, 0.753–0.841), and ResNet34 demonstrated a strong AUC of 0.92 (95% CI, 0.889–0.955). However, our state-of-the-art network stood out with the highest AUC of 0.995 (95% CI, 0.990–0.999), indicating its exceptional discriminative ability. Moreover, we assessed precision, recall, and F-measure metrics for both Stage III and Stage IV cases. For Stage III, DenseNet achieved precision, recall, and F-measure of 71%, 100%, and 83%, respectively. GoogleNet scored 84%, 60%, and 70%, while ResNet34 excelled with 88%, 98%, and 93%. Remarkably, our state-of-the-art network achieved precision, recall, and F-measure of 92%, 95%, and 94%. For Stage IV, DenseNet achieved precision, recall, and F measure of 100%, 60%, and 75%, respectively. GoogleNet scored 69%, 88%, and 77%, ResNet34 exhibited 97%, 87%, and 92%, and our state-of-the-art network maintained strong performance with 95%, 92%, and 93%. In conclusion, our study has demonstrated the effectiveness of various 3D neural network architectures in classifying the stages of OC in patients. While GoogleNet, DenseNet, and ResNet34 performed well they didn't quite achieve the desired accuracy and AUC levels for this task. However, our innovative cutting-edge network emerged as the winner by outperforming all models in terms of accuracy and AUC. This makes it the suitable choice for diagnosing this condition. Additionally, when examining precision, recall, and F-measure metrics our cutting-edge network demonstrated performance in distinguishing between Stage III and Stage IV cases. This ensures a level of confidence in the process. These findings highlight the importance of developing customized learning models for medical tasks. Our cutting edge network demonstrates its potential as a valuable tool in medicine. Future research can explore how to enhance and apply this network to improve healthcare outcomes. Conclusion:

Keywords: Ovarian cancer, Staging, Deep learning