Medical Image Captioning in Kurdish Language Using AI

الملخص

The diagnosis of thoracic disorders such as cardiomegaly, effusion, and pneumothorax depends heavily on medical imaging, especially chest X-rays. However, skilled radiologists are needed to interpret these images, and their shortage in many areas causes delays in diagnosis. Recent developments in deep learning have shown promise in automating image interpretation through caption generation; however, the majority of current systems lack disease-specific contextual understanding and are trained on English-language datasets, which limits their clinical utility and multilingual applicability. The disease-aware medical image captioning model developed for the Kurdish (Sorani) language is presented in this research. It produces textual descriptions for chest X-ray pictures that are both semantically rich and clinically correct. The suggested model uses a Bahdanau attention-equipped LSTM decoder to guide caption generation and incorporates disease-specific class embeddings to extract spatial features using CheXNet (DenseNet-121), which was pretrained on ChestX-ray14. To train and assess the system, a new Kurdish dataset was created, which included 7,775 chest X-ray pictures from nine different diagnostic categories together with carefully chosen Kurdish descriptions. A BLEU-4 score is 0.4015, a ROUGE-L score is 0.8722, and a BERTScore F1 is 0.9151. The model beat the current benchmarks when performance was assessed using BLEU, ROUGE-L, BERTScore, and CHRF++. These findings show how well the system produces linguistically coherent, disease-relevant captions in a language environment that has limited resources. This work shows a high-quality Kurdish dataset to help future research in multilingual clinical AI applications, in addition to developing the state of medical image captioning for presented languages that are rare.