Automatic disease detection and medical report generation from medical images using deep learning methods

Chest X-ray images play a critical role in the diagnosis of many diseases, but analyzing and interpreting images using traditional methods is a time-consuming process that requires specialized personnel and carries a high risk of human error. The aim of this thesis is to develop deep learning-based, high-performance, low-cost, and flexible models for the diagnosis and reporting of multi-labeled diseases from chest X-ray images. The study uses the ChestX-Ray14 dataset, which contains 112,104 labeled chest X-ray images, the Indiana University chest X-ray dataset consisting of chest X-ray images and reports, the ROCOv2 dataset used in medical image captioning studies, and a Turkish medical reporting dataset created from chest X-ray images and reports with ethical committee approval. Within the scope of the thesis, five different applications were developed to study the reporting of medical images. In the first application, EfficientNet and a coordinate attention mechanism were combined to develop a low-cost, high-performance model, achieving an AUC of 0.8309 in the classification of 14 diseases. In the second application, a hybrid model combining ConvNeXt and GRU architectures in an encoder-decoder infrastructure called G-CNX was used to generate paragraph-level reports, achieving the highest reporting score in the literature of 0.6544 on the Bleu-1 evaluation metric. In the third study, the model using the distillation technique called DeiTGPT attracted attention with its 66% time efficiency. In the fourth application, the Vi-Ba architecture, which draws power from transformer architectures, achieved a Rouge score of 0.274 and stood out with its flexibility. In the final application, the Model-SEY architecture, which focuses on Turkish medical report generation, achieved a score of 0.6412 on the Bleu-1 metric, producing effective results both technically and linguistically. The models proposed in this thesis can increase the efficiency of medical reporting, reduce the workload of specialist doctors, improve the quality of diagnostic processes, and prevent potential human errors.