Histotyping is a recent initiative, originating at the Institute in parallel with the recent wave of machine learning methods. Encouraged by promising initial results and great potential, we anticipate that Histotyping will play a central role at the Institute in the years to come.
Histotyping describes methods that make predictions about cancer patients by directly analysing digitally scanned histological sections of the tumour.
The methods are fully automatic and applied to digital scans without human intervention. Currently, we use methods based on deep learning with artificial neural networks, an appropriate choice given the strong performance of deep neural networks in image analysis. We train the networks to make predictions directly from the input images, allowing the method itself to learn which image features are important for the specific task.
The first results from the first Histotyping project came in the spring of 2019, following about two years of development. The goal was to develop a method that directly predicts the outcome of patients with early-stage colorectal cancer (CRC) using scans of H&E- stained tissue sections from the resected tumour. We evaluated the performance of the method, the DoMore-v1-CRC marker, on a predetermined validation set. The method achieved a hazard ratio of 3.84 in univariable analysis between patients classified as poor prognosis and patients classified as good prognosis; the CRC Histotyping method and results are published in The Lancets february 2020 edition.
As deep learning methods show strong performance in image classification, we will likely continue to use them in Histotyping. However, they present some conceptual challenges. We know that the method classifies an image based on the evidence in the training set, which suggests that it is more similar to other images with this class than to images with a different class. Gaining knowledge beyond this is very difficult due to the enormous complexity of the methods.
The DoMore-v1-CRC marker was trained and validated using a single tissue section per patient, but even within this small region, there is great variability. Some regions have normal tissue and are completely benign, while others have cancerous tissue expressing different features. The DoMore-v1-CRC marker attempted to remedy this intratumour heterogeneity using multiple instance learning, and the results are encouraging, albeit with room for improvement. Tumour heterogeneity is also much more prominent in some cancer types, like lung and prostate cancer which are known to be frequently multiclonal, which may warrant adaptation of the method
A digital image of a section scanned at the highest resolution provided by most scanner vendors, is typically about 100 000 x 100 000 pixels. This resolution is multiple orders of magnitude larger than images that can be directly analysed with current conventional deep learning classification methods on consumer-grade hardware. This makes input data size a challenge in terms of the methods we develop, which software we use to implement them in, and which hardware we run them on.
As there are still challenges to overcome, we will continue to scrutinise the current method to identify potential improvements. Methodological improvements, coupled with advances in software and hardware, make us excited about future results.
Other cancer types
We are confident that the same deep learning network can be trained to perform well on most other cancer types. For lung cancer, we already have two datasets with detailed clinical data, one containing more than 800 patients resected for primary lung cancer (Oslo cohort) and 6000 H&E-stained tumour tissue sections (one routine and one in-house section from each tumour tissue block), while the other containing more than 500 patients and sections (Tromsø cohort). We are currently in the process of training the deep learning network using these data but want additional datasets for validation
- We have for prostate cancer currently five available cohorts that may be included in a Histotyping design to test DoMore Network V1 as a prognostic marker in prostate cancer.
- For endometrial cancer have we several different cohorts.
The article "Deep learning for prediction of colorectal cancer outcome: a discovery and validation study" was published in the world's leading medical journal for global health; "The Lancet", on February 1st 2020. The assay has been extensively evaluated in large, independent patient populations, correlates with and outperforms established molecular and morphological prognostic markers, and gives consistent results across tumor and nodal stage. As a supplement to the article a video explaining the DoMore!-CRC-v1-marker was provided by the ICGI (see above). In addition a video providing a quick introduction to Histotyping was produced.
For more around histotyping, please see the Norwegian Research Council Lighthouse Project: DoMore! website.
This text was last modified: 13.02.2020