Recognition of honeycomb pattern in high-resolution computed tomography in patients with idiopathic pulmonary fibrosis using the deep learning model you only look once (YOLO)

Purpose: Medical imaging is becoming increasingly important for disease diagnosis, prognosis and treatment planning in medicine. Diagnosis of idiopathic pulmonary fibrosis using high-resolution computed tomography (HRCT) is a difficult task and can have high interobserver variability even in experienced radiologists. Deep learning is a quantitative approach that measures and extracts high-dimensional imaging features to aid in clinical decision making. It has shown great potential for use in the diagnosis, treatment and follow-up of many diseases using medical imaging. Idiopathic pulmonary fibrosis (IPF) is a disease with a high mortality rate and there are still some difficulties in diagnosing the disease. In our study, it was aimed to help in early recognition and diagnosis of the honeycomb pattern observed in HRCT with the deep learning method You Only Look Once (YOLO). Materials and Methods: In our study, 132 cross-sectional HRCT images containing honeycomb images belonging to 66 patients were used. Honeycomb labeling was performed on the images using Roboflow, a web-based platform designed to be used in labeling images for object detection, classification, and segmentation. The honeycomb pattern in the labeled HRCT images was introduced to the deep learning model You Only Look Once-V9 (YOLO-V9). Of the 132 cross-sectional HRCT images; 92 (%70) were used as training set data, 26 (%20) as validation set data, and 14 (%10) as test set data. 100 iterations (epochs) were studied on the images. In the training set data that we labeled and gave to the YOLO model, the model was first allowed to recognize the honeycomb pattern. With the validation set data, the YOLO model compared the real honeycomb images with the honeycomb images marked by the model thanks to the program it created. Then, using the test set data, the YOLO model estimated the probability of honeycombs on images it had never seen. Results: It is observed that the losses are generally low and stable on the training set data. In particular, the box_loss and dfl_loss averages are observed around 1.2. The losses on the validation set data are similarly around 1.7, indicating that the performance of the model on the validation data is similar to the training data. The loss rates are observed at a low level, which shows that the model is optimized step by step during the training process. The precision average is calculated as 0.538, and the recall average is 0.483, indicating that the model detects objects with reasonable accuracy. The mAP_0.5 average is calculated as 0.469 and the mAP_0.5:0.95 average is 0.246, indicating that the overall performance of the model is good, but the performance decreases as the IoU threshold increases. It has been shown that the YOLO-V9 model detects the honeycomb pattern with 81.3% accuracy, 68.3% sensitivity and 69.4% precision, and it can be said that the overall performance of the model is good. Conclusion: In our study, the most up-to-date version of YOLO V9, was used. It has a high original value in terms of being the first study conducted with YOLO for this disease. It can be said that the YOLO-V9 model generally exhibits acceptable performance in detecting the honeycomb pattern. The data set can be increased to increase the model accuracy. Nevertheless, good performance was achieved with these small data sets. A larger data set is likely to have a transformative effect in the field of pulmonology diagnosis