Efficientnet-B0 Based Single Stage and Multi-Scale Object Detection Model for Localization of Low-Grade Gliomas and Detection of 1p/19q Codeletion Status

Öksüz C., Urhan O., Güllü M. K.

Artificial Intelligence Theory and Application, vol.1, no.2, pp.191-201, 2021 (Peer-Reviewed Journal)


Introduction-Objectives: In recent years, it has been revealed that the codeletion of the 1p/19q chromosomal arms in low-grade glioma (LGG) patients is an important biomarker. Accordingly, codeletion is highly correlated with a patient’s positive response to treatment (chemotherapy, radiotherapy) and longer survival. Therefore, it is important to determine 1p/19q codeletion to make the correct treatment planning. For this purpose, a low complexity, single-step deep learning-based method is proposed that localizes the tumor region and predicts the codeletion simultaneously. Materials-Methods: In the study, the lightweight EfficientNet-B0 model is transformed into a multi-scale single-stage object detector model by inspiring the YOLO model architecture. Then, the model is trained by using the transfer learning strategy. Besides, pre-trained models with low complexity, i.e. Tiny YOLO v2, Tiny YOLO v3, and Tiny YOLO v4, are fine-tuned. Results: All the models are cross-validated on a public data set consisting of 159 patients with LGG (codeletion: 102, no-deletion: 57). The data set includes biopsyproven samples. The model performance scores, namely the log-average miss rate and mean average precision (mAP@[0.5:0.95]), are 0.14 and 40.21%, respectively. The method predicts codeletion with 90.2% sensitivity in 67ms inference time. Conclusions: In the study, promising results are obtained in the prediction of 1p/19q codeletion status by a fully automatic method. The performance obtained shows that the proposed model can be used as an assistive diagnostic tool for experts. As the data sets are updated with more samples, the robustness of the model can be increased by improving the performance.