Motivation

In recent years, Artificial Intelligence (AI) has made tremendous advances in identifying diseases from radiology images. Convolutional Neural Networks (CNNs), a class of deep learning algorithm trained on large volumes of labelled radiological images, have led these advances. Various results has shown that CNNs improves the speed, accuracy and consistency of diagnosis [Liu et al., 2017].

However, the adoption of deep learning diagnostic system by healthcare practitioners is prevented by two major challenges: 1) interpreting the prediction outputs from a deep learning network is not trivial, and 2) Privacy of patient data is not guaranteed when using online services that provides deep learning models.

These two challenges may be part of the reasons why healthcare practitioners remain wary of using AI-driven diagnostic [Ribeiro et al., 2016].

A medical practitioner cannot fully trust the CNN network except it can explain its logic, semantically or visually. Earlier methods in machine learning are transparent in how they compute the predictions but deep learning models are not so.

Deep learning models automate the hand crafted feature engineering and hence no knowledge of how the predictions are computed. Diagnosing with CNN involves studying image regions that contribute most to prediction outputs at the pixel level. In interpretability, we expect the CNN to explain its logic at the object-part level.

Given an interpret-able CNN, previous work reveals the distribution of object parts that are memorized by the CNN for object classification Wang et al. [2020].

Typically, a CNN model is being deployed through a server client architecture which requires the data to be sent to the model online for prediction. Deep learning models are large in memory and computation. Hence, they need large computing power like GPUs that requires an existing remote servers.

To get predictions, doctors have to upload the patient radiological scan through the internet making it at risk of data privacy. What we do in this work instead, is to use solutions that make such models run locally on the web browser thereby solving the issue of privacy. This technique also solves the challenge faced in developing countries where access to internet could be expensive.

Goal

In addressing these two challenges earlier discussed, the goal of this project is to therefore build a locally run web based application for interpreting deep learning models on breast cancer diagnosis. The project has no interaction with internet neither does it collect data from user. It is locally run on the web browser of the end-user.

Proposal

In other to have a CNN model having some level of certainty in its prediction, we would use out of distribution prediction method as the first step in evaluating our CNN model. This checks for a image belonging to the data distribution or not. Leveraging on the baseline methods discussed in (Cao et al. [2020]), we aim to experiment with these methods using dataset on breast cancer diagnosis (Araújo et al.).

This is a classification problem where the model outputs a sample has belonging to one of; normal, benign, in situ carconima or invasive carcinoma. We aim to extend this decision by explaining why the model make such a decision. We would achieve such by leveraging works done in model interpretability and explainability in (Wang et al. [2020]). The system shows a graphical visualization of the model outputs and its interpretation.

The novelty of our proposal is that, we are building this AI system with privacy concerns in mind and also overcoming challenges faced in developing countries like access to internet. Previous work done in medical diagnosis using CNN are large models trained on very large datasets which in result take up huge memory space and computation.

Thereby making it extremely difficult to be used on edge devices. With the long term goal of deploying on a local web browser using tools like tensorflow.js, we would also put into consideration post training model quantization. Jacob et al. [2017] which reduces the model parameters and making it easily deploy-able on a low resource device.

Long-term vision

The long-term vision of this project is that our software system will help radiologists to effectively and efficiently diagnose breast cancer in hospitals and clinics across Ghana, Cameroon, and ultimately across Africa. Importantly, our interpretable models will help radiologists to better understand the predictions given by the model, and ultimately provide safer medical care to patients.

To scale our diagnosis system across Africa, we plan to open-source our research code. This will allow machine learning researchers in other African countries to train our machine learning models, in partnership with their own local hospitals and data, and deliver reliable diagnoses to broad populations of African people. Adding X-Rays, MRIs and other medical image modalities to our diagnostic suite of software is also part of the long term vision.

Personnel

  1. Jeremiah Fadugba, Core team member, project lead. M.Sc in Mathematical Sciences, African Institute for Mathematical Sciences (AIMS), Ghana. Has 3 years experience as Machine Learning Engineer.
  2. Moshood Olawale, Core team member. M.Sc in Mathematical Sciences, African Institute for Mathematical Sciences (AIMS), Cameroon. Has 1 year experience as a Machine learning Engineer.
  3. Conrad Tankou, Medical domain expert, external advisor. Doctor of Medicine, University of Yahounde, Cameroon. He is a medical professional and radiologist. He works on cancer diagnosis.
  4. Oluwayetunde Sanni, Team member, Software Engineer, MSc AI and Robotics, Sapienza University of Rome, Italy. More than 5 years experience in software engineering.
  5. Two Research Interns. We have budgeted to hire two research interns who will be responsible for evaluating machine learning models and developing our system.