A Computer-aided Diagnosis System for the Automated Detection of Prostate Cancer on Whole-Mount Histology Images
Honorable Mention - Imaging InformaticsJames, Peter, Monaco ; Rutgers Unversity;
Content:
We present an automated means for detecting prostate cancer from images of Hematoxylin and Eosin stained whole-mount histological sections. The system is tailored to operate at low resolution (0.01mm2 per pixel). At low resolution glands are the primary visible structures. Accordingly, they are classified as either cancerous or benign using two discriminating features: 1) gland area and 2) the proximity of cancerous/benign glands to other cancerous/benign glands. The malignant glands are then aggregated into regions indicating the spatial extent of cancer.
Technology:
A set of Hematoxylin and Eosin stained prostate whole-mount histological sections are digitized using a whole-slide digital scanner. These images are analyzed using a set of routines implemented with the MATLAB software package.
Design:
The cancer detection algorithm begins by segmenting the individual glands. Using these segmentations, the area of each gland is determined. A classifier then assigns a probability of cancer to each gland based on its area. Glands whose probabilities exceed an empirically chosen threshold are labeled as cancerous. The remaining glands are labeled benign. A Markov Random Field iteration refines these labels by encouraging neighboring glands to be labeled similarly. Finally, the cancerous glands are aggregated into contiguous regions.
Results:
Figure (a) shows a whole-mount histology section with the cancerous extent as determined by a pathologist roughly encircled in black. Figure (b) illustrates the result of gland segmentation (blue boundaries). The white box in (b) is magnified in (c). Figure (d) indicates the centroids (green dots) of those glands whose probabilities of cancer exceed the selected threshold. These labels are then refined by a Markov Random Field iteration, producing the labeling shown in (e). Figure (f) illustrates the cancerous extent as determined by our automated system (green) and by an expert pathologist (yellow). Our algorithm detects cancer with a sensitivity of .8670 and a specificity of .9524
Conclusion:
We have presented an automated system for the detection of prostate cancer in whole-mount histology images. The method is specifically designed to operate at low-resolutions where the primary visible structures are glands.
