MATLAB Biopsy Analysis

Image read-in

Each image within the training dataset is read in individually for analysis. The image to the left is a real biopsy image of a prostate cancer tumor, obtained from an online database. Note, this image is for demonstration and was not included in the aforementioned 60,000 image dataset

Image enhancement

Each image is first enhanced in order to better collect data later on. This is accomplished through a complex process called contrast-limited adaptive histogram equalization (CLAHE) after grayscaling, which reduces noise and improves contrast relative to nearby image regions.

Image segmentation

Now, the program isolates only the nuclei in the image. Color k-means clustering is used first to perform simple isolation, then a watershed transformation which references local extended minima separates any clusters.

Data collection

Next, the program calculates characteristic data on each of the nuclei identified: radius, texture, perimeter, area, smoothness, and compactness. Following this, higher-level statistics are collected on these data in order to compile each image in to a single row of interpretable data for the machine learning program: mean, median, max, min, Q1, Q3, standard deviation, and number of nuclei outside of two standard deviations. Pictured left is a snippet of data gathered from 100 images.

Machine learning model fit

A K-Nearest Neighbor machine learning model is fit using both the gathered data from the training set images and their associated classifications. This model is saved for future use.

Novel image classification

Using the image analysis process described above, the trained machine learning model makes an "educated guess" as to whether a novel image is cancerous or non-cancerous. The program is currently set up to run through many images and check accuracy rather than to do single images, but simple adjustments can be made to accommodate this method.