Abstract
BACKGROUND: Pneumonia can be anatomically classified into lobar, lobular, and interstitial types, with each type associated with different pathogens. Utilizing artificial intelligence (AI) to determine the anatomical classifications of pneumonia and assist in refining the differential diagnosis may offer a more viable and clinically relevant solution. This study aimed to develop a multi-classification model capable of identifying the occurrence of pneumonia in patients by utilizing case-specific computed tomography (CT) information, categorizing the pneumonia type (lobar, lobular, and interstitial pneumonia), and performing segmentation of the associated lesions. METHODS: A total of 61 lobar pneumonia patients, 60 lobular pneumonia patients, and 60 interstitial pneumonia patients were consecutively enrolled at our local hospital from June 2020 and May 2022. All selected cases were divided into a training cohort (n=135) and an independent testing cohort (n=46). To generate the ground truth labels for the training process, manual segmentation and labeling were performed by three junior radiologists. Subsequently, the segmentations were manually reviewed and edited by a senior radiologist. AI models were developed to automatically segment the infected lung regions and classify the pneumonia. The accuracy of pneumonia lesion segmentation was analyzed and evaluated using the Dice coefficient. Receiver operating characteristic curves were plotted, and the area under the curve (AUC), accuracy, precision, sensitivity, and specificity were calculated to assess the efficacy of pneumonia classification. RESULTS: Our AI model achieved a Dice coefficient of 0.743 [95% confidence interval (CI): 0.657-0.826] for lesion segmentation in the training set and 0.723 (95% CI: 0.602-0.845) in the test set. In the test set, our model achieved an accuracy of 0.927 (95% CI: 0.876-0.978), precision of 0.889 (95% CI: 0.827-0.951), sensitivity of 0.889 (95% CI: 0.827-0.951), specificity of 0.946 (95% CI: 0.902-0.990), and AUC of 0.989 (95% CI: 0.969-1.000) for pneumonia classification. We trained the model using labels annotated by senior physicians and compared it to a model trained using labels annotated by junior physicians. The Dice coefficient of the model's segmentation improved by 0.014, increasing from 0.709 (95% CI: 0.589-0.830) to 0.723 (95% CI: 0.602-0.845), and the AUC improved by 0.042, rising from 0.947 to 0.989. CONCLUSIONS: Our study presents a robust multi-task learning model with substantial promise in enhancing the segmentation and classification of pneumonia in medical imaging.