Anant Madabhushi

Ph.D.

Emory University

Professor

PLCO (Learn more about this study)

PLCOI-917

Feb 16, 2022

Computerized image analysis of morphology, texture on H&E slides of colon cancer patients

Our project focuses on using computational analysis of morphological, textural patterns in pathology slides for predicting tumor biology, clinical behaviour, and treatment response. We are particularly interested in identifying Stage II colon cancers which are more aggressive and hence would benefit from adjuvant chemotherapy over those less aggressive colon cancers in which curative surgery is sufficient, and more aggressive Stage III CCs who could benefit from the escalation of therapy over standard therapy. There has been much progress on our side in using interpretable pathology image features to develop machine learning models for prognosis and predicting higher risk cancers specific for Colon cancer. In a recent work, computationally extracted spatial architecture features of tumor infiltrating lymphocytes TILs was proven prognostic in stratifying patients according to their risk for progression. Around 1036 features related to spatial architecture and density of TILs were extracted from each patient from a set of 120 patients (50 for training and 70 for validation). The resulted model was able to accurately predict higher incidence of progression in patients in the high-risk group (HR = 3.76, 95% CI 1.3-10.9, p-value=0.0053, c-index=0.687) in the validation set. In another work, we were able to distinguish Stage II from Stage IV CCs based on nuclear morphometric features. The initial discovery set was of 100 Stage II CCs with no evidence of recurrence for a minimum of 5 years and 100 Stage IV CCs with liver or lung metastases at presentation. The trained model was able to differentiate low and high-risk Stage II and Stage IV CC patients in terms of DFS (HR=2.20, CI= 1.46-8.770, p=0.006). Also, features quantifying the degree of disorder of collagen fiber orientations were also studied in this project and were also strongly associated with KRAS mutational status in Stage II and Stage IV CC. We also developed in our lab a tumor budding detection model which achieved high accuracies in detection in Oral cavity, and features extracted using the model was prognostic for overall survival. There are plans on implementing similar methods in Colon cancer.
As illustrated earlier, many computationally derived H&E image biomarkers have been studied and proven prognostic in our group. Also, we have deep learning models trained with high accuracies for segmenting biomarkers such as tumor infiltrating lymphocytes, Collagen fibre, and tumor budding. We also have experience in setting up pipelines which can extract nuclear shape, orientation, architecture, and texture features. Getting access to your dataset can help validate our findings and train more robust prediction models.