Register for free to
view the full innovation profile.
A deep learning-based algorithm applied to H&E digitised slide images of tumour tissue to stratify CRC patients into low- or high-risk of relapse of the cancer after surgery.
About
Adjuvant chemotherapy treatment is the standard of care for patients with Stage II/III colorectal cancer (‘CRC’). However only about 20% of patients will benefit from this treatment. The traditional pathological biomarkers of T, N, V and L status do not adequately stratify the stage II/III CRC patients and therefore do not give detailed enough information to allow clinicians to make coherent individualised decisions about the type and duration of adjuvant chemotherapy. This has led to quite different patterns of clinical treatment between countries, between different cancer centres in the same country, and in some cases, divergence between different clinicians practising in the same cancer centre. OCB have developed OncoProg(Patent Number: UK 1504569.3, PCT/EP2016/055102, US-2018-0075598-A1, Japan 2017/567536), an in vitro diagnostic (IVD) medical device that uses digital pathology to determine the DNA content (ploidy) and tumour microenvironment (stroma content) of formalin-fixed paraffin embedded (FFPE) tissue from patients with Stage II / T3N1 colorectal cancer. These two biomarkers are combined to stratify patients into categories informing on risk of recurrence, which allows clinicians to make more informed decisions on any adjuvant chemotherapy treatment strategy. The device has been CE marked and has just entered the UK’s healthcare market (NHS and private sector), which is showing early signs of commercial success.
This innovation is the partial output of the continuous development of the OncoProg product. It is based on our existing knowledge in this field and is about implementing a brand new novel biomarker for same patient group but using artificial intelligence (AI) deep learning (DP) technologies on whole slide images (WSIs) of H&E (Haematoxylin and Eosin) stained tumour tissues.
The following image sets are used in the development:
1) NCT-CRC-HE-100K: a public H&E-stained image dataset of colorectal cancer comprised of 100,000 image patches each with a resolution of 224 × 224 pixels.
2) An image set of fifty-four stage III colon cancers with more than 36 months follow-up from TCGA-COAD (https://portal.gdc.cancer.gov/projects/TCGA-COAD).
3) An image set of 168 WSI images from the 168 surgical specimens of the CRC III patients in West China Hospital (WCH), Sichuan University. OCB is collaborating with WCH for in this innovation.
4) An image set of 1047 WSI images from an open-label, randomised, controlled QUASAR 2 trial, which was done at 170 hospitals in seven countries for histologically proven stage III or high-risk stage II colorectal cancer.
The innovation includes the following content:
1) Automatic tumour area detection
The standard operations in a pathology lab involve experienced pathologists to manually mark the tumour region (region of interest, ROI) of tissue slides before further quantisation examination is made. A CNN (Convolutional neural network) data model within the tensor-flow objection detection framework was developed for automatically detecting the ROIs on the H&E WSI images. The 1047 QUASAR 2 images were used to train a selection of pre-trained models including MobileNetV2, ResNet102 to build the data model.
2) WSI Tile Categorisation
Each WSI is partitioned into multiple non-overlapping image regions called tiles with a resolution of 224 × 224 pixels. A data model was trained for categorising the tile image into one of the following eight categories: adipose tissue (ADI), debris (DEB), lymphocytes (LYM), mucus (MUC), smooth muscle (MUS), normal colon mucosa (NORM), stroma (STR), and colorectal adenocarcinoma (TUM). A selection of pre-trained CNNs including VGG19, ResNet50, ResNet102, InceptionV3, InceptionResNetV2 were tried during the development. The model of the best performance achieved an accuracy of 99%.
3) Gradient Boosting Classifier for predicting cancer recurrence
After the WSI tile partitioning and categorising, the proportions of each tissue tile category (8 categories) in each whole-slide were counted and were employed as features in a machine classifier for the prediction of recurrence of the CRC. The proportions of five tissue categories (DEB, LYM, MUC, STR and TUM, discarding the normal components - ADI, MUS, and NORM) are used to generate 45 new parameters (morphologic parameters) for the classifier by combing each two tissue categories, such as DEB_proportion/LYM_proportion, DEB_proportion/MUC _proportion etc., which got 15 continuous variable parameters (5 original proportion, and 10 combined ratios). Each case was assigned to “
Key Benefits
OCB’s CE-marked OncoProg product is currently the only digital pathology solution available on the market. It uses a combination of two biomarkers, ploidy and stroma, which requires two separate chemical preparations for the tissues and one of them needs a special nuclei digestion and a Feulgen-Schiff staining processes, therefore leads to a higher cost. The new AI-based prognostic technology only requires H&E slide images, which are the output of the typical pathological process and ubiquitously available in the digital pathology lab. The application of this innovation will remove the need of nuclei wet preparation process and will therefore significantly save the lab setup cost, reduce the test processing time, and significantly reduce the cost of assay.
Applications
CRC is the second most common in women and third most common cancer in men and there are 1.4 million newly diagnosed patients each year worldwide. The annual costs of CRC treatments in EU is over €13bn (10% of total cancer related costs). According to the statistical data of 2014-16, there were 42,042 new cases in average each year the UK. Around 50% of CRC patients are at Stage II/III. OCB’s current OncoProg assay was initially developed for CRCII/III patients and then has been expanded to patients of prostate, ovarian and gastric cancers. There were no other similar tests in use in a clinical setting. Genomic Health had previously developed a gene profiling technique (Oncotype Dx Colon) for determining a patient’s risk of recurrence but due to poor performance this test is not commercially available in the UK. Other companies in the digital pathology area are developing tools for automated identification of tumour margins and automated tumour staging tools. Microsatellite instability testing is the only recognised tool available to clinicians, but this is not widely available in the UK. The clinical risk report provided by OncoProg to physicians details the relative risks of cancer recurrence with confidence intervals from the clinical validation data set. This information will assist the clinician and patient in making their decision on whether or not to accept adjuvant chemotherapy and which type of chemotherapy (single agent relatively low toxicity or combination chemotherapy with an associated higher incidence of side effects). For example, patients at high risk of relapse may be advised by the treating clinician to have multi-agent chemotherapy, those of intermediate risk (all groups not identified as High or Low risk) may be advised to receive single agent fluoropyrimidine treatment and those with low risk of relapse may be advised that observation would be the most appropriate ‘intervention’.