ABSTRACT
BACKGROUND AND OBJECTIVE: Medical imaging techniques are widely employed in disease diagnosis and treatment. A readily available medical report can be a useful tool in assisting an expert for investigating the patient's health. A radiologist can benefit from an automatic medical image to radiological report translation system while preparing a final report. Previous attempts on automatic medical report generation task includes image captioning algorithms without taking domain-specific visual and textual contents into account, thus arises the question about credibility of generated report. METHODS: In this work, a novel Adaptive Multilevel Multi-Attention (AMLMA) approach is proposed by offering domain-specific visual-textual knowledge to generate a thorough and believable radiological report for any view of a human chest X-ray image. The proposed approach leverages the encoder-decoder framework incorporated with multiple adaptive attention mechanisms. The potential of a convolutional neural network (CNN) with residual attention module (RAM) is demonstrated as a strong visual encoder for multi-label abnormality detection. The multilevel visual features (local and global) are extracted from proposed visual encoder to retrieve regional-level and abstract-level radiology-based semantic information. The Word2Vec and FastText word embeddings are trained on medical reports to acquire radiological knowledge and further used as textual encoders, feeding as input to Bi-directional Long Short Term Memory (Bi-LSTM) network to learn the co-relationship between medical terminologies in radiological reports. The AMLMA employs a weighted multilevel association of adaptive visual-semantic attention and visual-based linguistic attention mechanisms. This association of adaptive attention is exploited as a decoder and produces significant improvements in the report generation task. RESULTS: The proposed approach is evaluated on a publicly available Indiana University chest X-ray (IU-CXR) dataset. The CNN with RAM shows the significant improvement in recall (0.4423), precision (0.1803) and F1-score (0.2551) for prediction of multiple abnormalities in X-ray image. The results of language generation metrics for proposed variants were acquired using the COCO-caption evaluation Application Program Interface (API). The trained embeddings with AMLMA model generates the convincing radiology report and outperform state-of-the-art (SOTA) approaches with high evaluation metrics scores for Bleu-4 (0.172), Meteor (0.247), Rouge_L (0.376) and CIDEr (0.381). In addition, a new "Unique Index" (UI) statistic is introduced to highlight the model's ability for generating unique reports. CONCLUSION: The overall architecture aids to the understanding of various X-ray image views and generating the relevant normal and abnormal radiography statements. The proposed model is emphasized on multi-level visual-textual knowledge with adaptive attention mechanism to balance visual and linguistic information for the generation of admissible radiology report.
