Open AccessDevelopment of Deep Learning-Based Automatic Detection Algorithm for Adrenal Nodules on Contrast-Enhanced Abdominal CT Scans
Author(s) -
Chang Ho Ahn,
Tae Woo Kim,
Kyungmin Jo,
Sung Hye Kong,
Jinhee Kim,
Taesung Kim,
Chan Soo Shin,
Jaegul Choo,
Jung Hee Kim
Publication year - 2021
Publication title -
journal of the endocrine society
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.046
H-Index - 20
ISSN - 2472-1972
DOI - 10.1210/jendso/bvab048.166
Subject(s) - nodule (geology) , medicine , radiology , receiver operating characteristic , algorithm , adrenal gland , false positive paradox , asymptomatic , nuclear medicine , artificial intelligence , pathology , computer science , biology , paleontology
Objective: Adrenal nodules are often incidentally detected on abdominal computed tomography (CT) scans due to their asymptomatic nature. We aimed to develop an automatic detection program for adrenal nodules on abdominal CT scans using deep learning algorithms. Methods: We retrospectively analyzed abdominal CT scans performed at two university-affiliated hospitals (n = 483 and n = 514, respectively) from 2006 to 2019. This dataset was randomly divided into training set (181 CTs without adrenal nodule and 362 CTs with adrenal nodule) and test set (291 CTs without adrenal nodule and 163 CTs with adrenal nodule). All CT scans were contrast-enhanced and the phase with the highest contrast between adrenal gland and adjacent normal tissues was selected for multi-phase CT. The core algorithm of our deep learning algorithm for adrenal nodule (DLAAN) was MULAN (Multitask Universal Lesion Analysis Network) algorithm whose backbone was a convolutional neural network. DLAAN was composed of two stages. The first stage was to detect the CT slice where normal adrenal gland or adrenal nodule were located. The second stage was for fine localization of adrenal nodule on the corresponding CT slice. The performance of DLAAN was evaluated using the area under the receiver operating characteristic curve (AUROC) for patient-level classification and free-response ROC for nodule-level localization. The figure of merit for free-response ROC was calculated as an average sensitivity when 0.5, 1, 2, and 4 false positives per slice were allowed. Results: The AUROC of DLAAN was 0.927 (95% confidence interval: 0.900–0.955). With a threshold probability of 0.9, the sensitivity and specificity were 86.5% and 89.0%, respectively. When left and right adrenal nodules were analyzed separately, the AUROC was 0.910 for left adrenal nodule and 0.957 for right adrenal nodule, respectively. The accuracy of DLAAN according to the size of adrenal nodule was 0.890, 0.734, 0.981, 1.00 and 1.00 for no adrenal nodule, adrenal nodule sized 1–2 cm, 2–3 cm, 3–4 cm and > 4 cm, respectively. The performance of DLAAN for the localization of adrenal nodule which was estimated by average sensitivity was 0.812. The number of CTs with at least one false positive nodule was 93/454 (20.5%). Conclusion: Our proof of concept study of deep learning-based automatic detection of adrenal nodule on contrast-enhanced abdominal CT scans showed high accuracy for both the classification of patients with or without adrenal nodule and the localization of adrenal nodule, although the performance of the algorithm decreased for small sized adrenal nodules. External validation with different CT settings and patient population is needed to assess the generalizability of our algorithm.