An image segmentation procedure that uses a 4-layered U-Net Architecture to segment tumours from Ultrasound images of the breast.

tumourbw maskbw

Note: The BREAST TUMOUR SEGMENTATION notebook above contains an idepth explanation of the proposed procedure. This the a dumped down version of the contents in the notebook

• The data used in this project contains 210 images of malignant tumours, 437 images of benign tumours and 133 images of normal Ultrasounds creating a total of 780 images
• Each image in the dataset contains varying image sizes ranging from 200x150 to 300x300.
• A Ultrasound can have more than one mask depending on the number of tumours present in the image
• each Ultrasound image is saved as a jpg file instead of the usual dicom format
• The dataset is available here

• The images are first resized to a dimension 224x224
• Each image is then coverted to grayscale for simplicity
• The masks are binarized. That is, each pixel is changed to either be 0 or 1
• The Ultrasounds are the normalised between 0 and 1

A 4-layered generic U-Net Architecture is trained to effectively segment the tumours. The various levels have fiilter sizes of 32, 64, 128, 256 respectively

• The Model achieved a training dice score of 92%
• The Model achieved a testing dice score of 74%

tumour mask segmented

• The model is really good as identifying isolated/exposed tumours
• The model is good at getting the shape of the the tumour but lacks the fine outlines on the prediction
• The model fails to identify tumours partially hidden by tissue. This is evident in plot_test(11)
• The model also mistakens writings or drawings as an indication of a tumour. This is evident in plot_test (29)
• The model to generally overfits to the training images