Enhancing Chest X-ray Analysis using Encoder-Decoder with GRU for Report Generation

Despite the strides made in deep learning, the development of a reliable framework tailored explicitly for generating chest X-ray reports remains a persistent challenge. The study addresses critical gap by introducing an innovative architecture crafted to extract intricate features from chest X-ray images and articulate comprehensive diagnostic information within generated reports. The approach includes the utilization of DenseNet121 as the image feature extractor, leveraging its pre-trained knowledge from ImageNet. The strategic choice confers a pivotal advantage by transferring learned feature representations to the domain of chest X-ray analysis. Fine-tuning DenseNet121 through end-to-end training not only harnesses these pre-existing features but also tailors them specifically for the nuanced task of report generation. The approach circumvents overfitting and the prolonged training typically associated with training a Convolutional Neural Network (CNN) from scratch. An embedded layer augments the architecture by fostering semantic relationships between words, complementing the two-stage encoder-decoder structure. The design ensures a clear specialization: the CNN encoder excels in image feature extraction, while the GRU decoder focuses on sequence generation. The holistic end-to-end training methodology enables simultaneous learning of optimal feature representations and the intricate mapping between images and reports, representing a significant leap towards more effective and detailed chest X-ray report generation. Most importantly, the suggested architecture scores an astounding 87.55% accuracy rate, indicating not only a good BLEU score but also remarkable precision in text generation and diagnostic knowledge. It validates the model's effectiveness and highlights its potential to significantly improve the diagnostic accuracy and text generation quality of chest X-ray reporting.