This work focuses on medical image processing, specifically on liver CT scans and segmentation masks. We start by loading .nii files, extracting pixel data, and returning it as numpy arrays. This process ensures correct data orientation for further processing. We verify data loading by sampling scans and masks, checking their shapes and paths. The code then introduces custom classes and functions for preprocessing, such as TensorCTScan for grayscale display and histogram-based normalization functions like freqhist_bins and hist_scaled. These functions create multiple channels for various contrast settings, essential for deep learning models. We save processed images in RGB format for compatibility. Our custom deep learning model combines Attention U-Net with pre-trained models, DenseNet121 and InceptionV3. The architecture uses attention gates for refined segmentation and a transformer block for feature refinement, leading to precise pixel-level boundaries. The model is optimized for high accuracy in liver tumor identification.

Keywords: Liver CT scans, segmentation masks, medical image processing, Attention U-Net, DenseNet121, InceptionV3, deep learning, data preprocessing, liver tumor identification.