Automated Thyroid Ultrasound Classification Using Cnn

This work focuses on developing an end-to-end approach in automatically classifying thyroid ultrasound images by using a compact convolutional neural network and metadata-driven labelling. The processing starts with unpacking an archive of ultrasound images together with XMLbased annotation files into the working directory. From such XML records, the case identifiers and TIRADS scores were extracted, then converted to simple binary labels that reflect whether an image does or does not show thyroid pathology. After organizing the images in class-specific directories, a TensorFlow data pipeline was built, which handled some essential preprocessing steps, including resizing, intensity normalization, caching, and prefetching to ensure efficient GPU utilization in training. The model used is a small, sequential CNN consisting of three convolution stages with 32, 64, and 128 filters, each followed by pooling, and ended with fully connected layers for prediction. For training, the model employed the Adam optimizer with a binary cross-entropy objective, while the performance was monitored on the separate validation subset. A simple inference module has also been included that preprocesses a given frame of ultrasound and displays the model's prediction alongside the visual input. In general, the design emphasizes practical aspects of implementation suitable for smaller datasets of medical images, while remaining reproducible and capitalizing on structured XML annotations for accurate label generation.