Table 1.
Specification of dataset.
Fig 1.
Overview of the proposed system.
Anatomical information and characteristics of the target nodules are extracted from CT images using image recognition modules. Text generation module integrates them to generate finding sentences. Reprinted under a CC BY license, with permission from Shingo Iwano, original copyright 2024.
Fig 2.
Creation of sparsely labeled data.
By analyzing the findings, the bronchopulmonary segments, where the lesion is located, can be found. Also, the coordinate information of the lesion is recorded as annotation data. This information is combined to create localized ground truth data of bronchopulmonary segments. Reprinted under a CC BY license, with permission from Shingo Iwano, original copyright 2024.
Table 2.
List of image findings of lung nodules.
Fig 3.
Size and characteristic distribution of lung nodules in the evaluation dataset.
Table 3.
Specification of evaluation dataset.
Fig 4.
The gold standard examples of the observer performance test dataset.
Reprinted from https://www.cancerimagingarchive.net/collection/rider-lung-ct/ under a CC BY license, with permission from Zhao, B., Schwartz, L. H., & Kris, M. G., original copyright 2015.
Fig 5.
Content and rank trend of the observer performance test dataset.
(A) contents that must be described, (B) contents that should be described, (C) any content, and (D) wrong content.
Fig 6.
Performance evaluation results and examples of bronchopulmonary segments prediction.
(a/b) Comparison of the exact match rate/partial match rate between ground truth and prediction results by AI with/without edge loss and two radiologists. * indicates p < 0.05. (c) Examples of prediction results. Different colors mean different segments. Reprinted under a CC BY license, with permission from Shingo Iwano, original copyright 2024.
Table 4.
Accuracy of bronchopulmonary segments prediction.
Fig 7.
Comparative evaluation results of lung nodule classification.
(a) Comparison of AUC for each class between AI with enlarged image input and one without. (b) AI and individual radiologist classification results. ** and * indicates p < 0.01 and p < 0.05, respectively.
Fig 8.
Performance evaluation results of lung nodule classification.
Table 5.
The result of observer performance test.
Fig 9.
Examples of cases including images and descriptions.
“Generated description” was generated by our system. “Reference description” was created by the co-authored radiologist from scratch (without our system). Both original sentences are written in Japanese and translated into English for explanation. The underlined word in generated description indicates the FP content of the generated description. The underlined word in reference description indicates the FN content of the generated description. Reprinted from https://www.cancerimagingarchive.net/collection/rider-lung-ct/ under a CC BY license, with permission from Zhao, B., Schwartz, L. H., & Kris, M. G., original copyright 2015.