Currently, radiologists face an excessive workload, which causes fatigue and, as a result, undesirable diagnosis errors. Decision support systems can help radiologists prioritize and make faster decisions. In this regard, medical content-based image retrieval systems that provide well-curated similar examples can be extremely useful. Nonetheless, most medical content-based image retrieval systems operate by locating the most similar image, which is different from locating the most similar image in terms of disease and severity.
Researchers have proposed an interpretability-driven and attention-driven medical image retrieval system. They ran tests on a large, publicly available dataset of chest radiographs with structured labels derived from free-text radiology reports. The methods were tested on two common conditions: pleural effusion and (possible) pneumonia. An experienced board-certified radiologist classified and ordered query/test and catalog images as ground truth for the evaluation. Additional radiologists provided their rankings for a thorough evaluation, allowing us to infer inter-rater variability and yield qualitative performance levels.
They have also quantitatively evaluated the proposed approaches based on ground-truth ranking by computing the normalized Discounted Cumulative Gain (nDCG). They discovered that the Interpretability-guided approach outperforms the other cutting-edge approaches and has the highest agreement with the most experienced radiologist. Furthermore, its performance is within the range of inter-rater variability observed.
Related Content: Small-Molecule Therapies – DNA Altering Diseases