Introduction

Introduction:

The application of imaging in the detection of lung diseases such as tuberculosis and pneumonia is increasing rapidly in the medical domain. The analysis of these medical images requires highly skilled medical professionals, especially trained and experienced radiologists, who need to undergo years of training to attain proficiency.

Given the rising demand for medical imaging and the time taken in the rigorous training processes, there is an acute shortage of experienced professionals (particularly radiologists) worldwide.
We have analyzed this gap in delivering Chest X-Ray analysis services to the patients, and propose an Artificial Intelligence (AI) based solution.

1.1 The Current Scenario

Chest X-ray analysis falls under the medical paradigm of radiology and requires radiologists to analyze the X-Ray images. Radiologists are required to examine the X-Ray in detail and write their observations in the form of impressions and findings. The findings are the detailed description whereas impressions are the crisp description for the same.
There are three major challenges that motivate us to further investigate the Chest X-Ray analysis using AI-based techniques.

We observe that due to the high demand for radiologists, the workload for a single radiologist tends to be significantly large and they are forced to look at their computer screens for several hours for analyzing these images. This problem is acute in highly populated regions like India and China. As a result, the radiologists tend to miss some of the radiological findings while reporting, which may lead to an erroneous diagnosis and treatment regimen.
Secondly, in the rural areas, the junior medical professionals lack sufficient practical training and tend to make judgmental errors while analyzing these X-rays leading to erroneous diagnosis and treatment for a patient.
Thirdly, we observe that nearly 80 percent of times, the Chest X-rays are normal. So, if we devise a solution that can at least distinguish a normal x-ray from an abnormal x-ray, we can reduce the workload of radiologists by 5 folds.

1.2 The Proposed Solution

The current paper presents a novel, deep learning-based architecture CARES for report generation which offers a) CNN feature extraction, b) ocular opacity, and anatomical knowledge infused Attention mechanism, c) LSTM based reconstruction module, d) Novel scoring mechanism - Radiological Finding Quality Index (RFQI) that helps in generating more grammatically and clinically accurate reports.

The key contributions of our work are:

We built an AI-based labeler on top of state-of-the-art CheXpert Labeler and evaluated its performance against the CheXpert labeler.
We propose a novel deep learning-based architecture (CARES) for report generation.
We propose a novel scoring mechanism (RFQI) that considers the exact radiological finding, localization, size/severity for each such term present in the report.
We perform multiple experiments on the in-house dataset to validate our results.