Abstract

ABSTRACT

Chest X-Rays are widely used in medicine for preliminary examination and treatment of patients. In most cases, this involves analyzing the X-Ray image, which can be a menial task. Moreover, the subsequent report can have human errors.

To overcome the potential for error and meet the growing demand for trained radiologists, the present paper looks at a novel deep learning-based architecture Convolution Attention-based sentence REconstruction and Scoring (CARES) for the identification and localization of radiological findings in a chest X-Ray image. The solution would also generate a clinically meaningful description to aid the radiologists, and deliver: a) CNN feature extraction, b) ocular opacity, and anatomical knowledge infused attention based graph, c) hierarchical 2-level LSTM based reconstruction module, d) pre-trained transformers along with clustering and scoring model that helps in generating more grammatically and clinically accurate reports.

In addition, the present paper further proposes a novel scoring mechanism: Radiological Finding Quality Index (RFQI) that considers the exact radiological finding, localization, size/severity for each such term present in the report. The AI-based labeler has been tested on an inhouse curated dataset for 5 categories and against the state-of-the-art CheXpert Labeler - i.e., Cardiomegaly (CM), Edema (ED), Consolidation (CD), Atelectasis (AT), and Pleural Effusion (PE). In comparison to the CheXpert labeler - (CM: 45794), (ED: 9272), (CD: 17891), (AT: 6095), and (PE: 24669), the proposed solution was able to detect (CM: 52843), (ED: 12532), (CD: 21742), (AT: 9643), and (PE: 30253) positive reports.

The solution’s multilabel image classification AUCs for the top 5 classes on 3 publicly available datasets are: Chexpert: 0.9412, PadChest: 0.9435, NIH: 0.8890 and for the inhouse curated dataset: 0.9420. The paper demonstrates that the proposed AI-based labeler outperforms the state-of-the-art CheXpert labeler on the inhouse curated dataset.