Content-based CT image retrieval system using deep learning: Preliminary assessment of its accuracy for classifying lesion patterns and retrieving similar cases among patients with diffuse lung diseases

Practical image retrieval systems must fully use image databases. We investigated the accuracy of our content-based computer tomography (CT) image retrieval system (CB-CTIRS) for classifying lesion patterns and retrieving similar cases in patients with diffuse lung diseases. The study included 503 individuals, with 328 having diffuse lung disease and 175 having normal chest CT scans. Among the former, we randomly selected ten scans that revealed one of five specific patterns [consolidation, ground-glass opacity (GGO), emphysema, honeycombing, or micronodules: two cases each]. Two radiologists separated the squares into six categories (five abnormal patterns and one normal pattern) to create a reference standard. Subsequently, each square was entered into the CB-CTIRS, and the F-score used to classify squares was determined. Next, we selected 15 cases (three per pattern) among the 503 cases, which served as the query cases. Three other radiologists graded the similarity between the retrieved and query cases using a 5-point grading system, where grade 5 = similar in both the opacity pattern and distribution and 1 = different therein. The F-score was 0.71 for consolidation, 0.63 for GGO, 0.74 for emphysema, 0.61 for honeycombing, 0.15 for micronodules, and 0.67 for normal lung. All three radiologists assigned grade 4 or 5 to 67.7% of retrieved cases with consolidation, emphysema, or honeycombing, and grade 2 or 3 to 67.7% of the retrieved cases with GGO or micronodules. The retrieval accuracy of CB-CTIRS is satisfactory for consolidation, emphysema, and honeycombing but not for GGO or micronodules.