[1452] Quantifying Gland Morphology for Computerized Prostate Cancer Detection and Gleason Grading.

Rachel E Sparks, Anant Madabhhushi. Rutgers University, Piscataway, NJ

Background: Observer variability between expert pathologists when assigning Gleason grade is high, particularly distinguishing Gleason patterns 3 from 4. Consequently there is need for decision support system that having extracted subtle, quantitative image descriptors can help distinguish intermediate Gleason patterns.
Gleason patterns differ subtlety in their in their gland morphology and nuclear structure, with pattern 3 glands being larger and more distinct compared to pattern 4 glands. Similarly benign prostate glands tend to be larger with distinct lumen and orderly nuclei Compared to malignant glands which tend to be small with indistinct lumen and irregular nuclei. In this work we present a novel scheme for extracting quantitative descriptors of gland morphology that can be used to distinguish (a) intermediate Gleason grade patterns, and (b) malignant and benign prostate areas.
Design: We present novel Explicit Shape Descriptors (ESDs) for modeling gland morphology. The method involves (a) modeling gland shape (via the medial axis transform as seen in Figure 1), (b) pairwise comparison between fitted shape models, (c) dimensionality reduction to obtain ESDs, and (d) a support vector machine classifier to distinguish (i) Gleason patterns, and (ii) benign and malignant areas.


Results: An expert pathologist assigned a Gleason pattern to 120 regions of interests (ROIs) from 58 prostate biopsy studies (serving as ground truth). 888 glands were selected from benign (24 ROIs, 93 glands), Pattern 3 (67 ROIs, 748 glands), and 4 (11 ROIs, 47 glands) ROIs.

Table 1
TaskMalignant v. BenignPattern 3 v. Other (Benign, Pattern 4)Pattern 4 v Other (Benign, Pattern 3)Pattern 3 v Pattern 4
AUC0.84±0.090.81±0.030.85±0.090.87±0.11
Area under the ROC curve (AUC).

Table 1 displays the AUC for four classification tasks, which is over 0.80 in all tasks.
Conclusions: We presented explicit shape descriptors for modeling gland morphology in order to distinguish (a) benign and malignant glands on needle cores, and (b) subtle differences between Gleason patterns 3 and 4. Future work will involve combining ESDs with architectural features for improving our Gleason grading classifier.
Category: Informatics

Tuesday, March 1, 2011 1:00 PM

Poster Session IV # 189, Tuesday Afternoon

 

Close Window