What Can AI Do for Pathology?
Accurate diagnosis and assessment of disease are fundamental in providing appropriate and timely treatment. However, interpretation of pathology images is subjective, qualitative, and manual. Machine learning can change this by learning to label images based on annotations by expert pathologists. Such automated analysis can detect disease, segment structures, classify slides, and quantify disease progression. By connecting histopathology images with clinical or genomic data, AI can even predict molecular properties and patient outcomes.
Automate labor-intensive or mundane processes
- Increase efficiency and repeatability
- Count mitoses, segment glands
- Characterize nuclei shape
- Improve productivity
- Find tumor in whole slide images
- Predict one imaging modality or stain from another
Discover and learn concepts beyond human capabilities
- Distinguish classes too complex for human experts
- Infer molecular biomarkers from H&E
- Predict patient outcome
How Does Machine Learning Work on Pathology?
Recent improvements in whole slide scanning systems, GPU computing, and deep learning make automated slide analysis well-equipped to solve new and challenging analysis tasks.
These learning methods are trained on labeled data. This could be anything from annotating many examples of mitosis, labeling tissue types, or categorizing a full slide or set of slides from a particular patient sample. The goal is then to learning a mapping from the input images to the desired output on training data. Then the same model can be applied to unseen data.
Rather than engineering specific features for each data type and task, modern machine learning methods learn the representation directly from the data. Deep learning is the most prevalent of these methods and is a way of learning a hierarchy of features where the higher level concepts are built on the lower level ones. Automatically learning these abstract features enables the system to learn complex functions mapping an input to an output without the need for hand-crafted features.
While deep learning is the most powerful learning technique for images today, some tasks are best-suited to other machine learning methods, for example on very small data sets or when interpretability is crucial.
Key Factors in the Success of a Pathology AI Project
Data availability and usage: Annotated data is essential as, generally, more data and finer-grained annotations make for a better model. However, pathology is a great example of where we can bend the rules a bit. Patient samples can be scarce and fine-grained annotations expensive, but the images are massive and can be used with weak labels. There may even be multiple sets of labels or modalities for each patient sample so that we can create a more robust model with multi-task learning. While data and annotation availability are critical, using what's available to its maximum potential can open up new possibilities.
Image variability: On small data sets it may be necessary to standardize the appearance of tissue using normalization. On larger data sets, however, capturing the full variability created by different staining techniques, intensities, and labs can make the model more robust. Simulating further variability in image appearance extends this technique and gives the appearance of a much larger data set.
Model initialization: Training a deep learning model from scratch is difficult and time-consuming. For most applications we employ transfer learning to make use of existing models from related domains. Unsupervised or self supervised learning may be the future, but transfer learning is very much the present state-of-the-art. It also provides a significant boost when data or annotations are limited.
How Can Pixel Scientia Labs Bring AI to Your Pathology Project?
We have 8 years experience in working with pathology images and have completed multiple projects with tasks ranging from segmentation to classification to outcome prediction. We are also experienced in integrating other forms of data such as genomics or clinical data.
A successful project is always a team effort between you and Pixel Scientia Labs. You know your data and your science best. We work to create new possibilities from that data.
When breaking new ground with science images we often start with a feasibility study. For tasks like predicting a molecular biomarker from H&E, we prototype the simplest possible solution to assess utility and get a lower bound on the classification accuracy.