Recently, Artificial Intelligence (AI) and Machine Learning (ML) have been demonstrating the potential to significantly improve the way radiology is practiced. One of the most promising applications of AI is cancer detection in mammography. This includes, but is not limited to, tumor detection, and region of interest analysis. To reach this goal, different ML techniques such as image processing, image segmentation, and image classification, can be applied. Once implemented, the AI will have a positive impact in improving the daily radiological workflow, increasing efficiency, accuracy, and reliability of services within the medical enterprise. Despite these potential advantages, AI faces several challenges on the path to becoming reality.
In this article, we explore the main challenges and limitations of AI in the medical imaging field and consider how NOVU will translate these potentially innovative technologies into readiness for clinical practice.
AI Challenges
The widespread implementation and integration of ML models into clinical practice require overcoming several challenges. Those challenges occur during the development and validation process of the model.
Limited volume and insufficient quality of data
Medical data is the building stone for developing efficient ML algorithms. It has been shown that the performance of any ML algorithm increases logarithmically based on volume of training data size [1]. Not only the volume but also the quality of the medical images impacts greatly the performance. Nowadays, publicly available medical imaging data for algorithmic training is not increased in size and suffers from bad quality. NOVU innovative software uses large and structured datasets. The dataset is fully annotated by professional radiologists and has minimal noise medical images which improves the accuracy and sensitivity of ML tasks.
Lack of robustness and reproducibility of ML models
Robustness and reproducibility of a ML model are two important factors for measuring the performance of the model [2]. If small changes to the input data have significant influence on the output, then the developed model is not robust. In addition, the developed model is not reproducible when different output results occur for the same scan. NOVU ensures the robustness and reproducibility by using large diverse datasets for training and validation.
Lack of generalizability of ML models
Most of the developed ML models are based on training one dataset from one source. The output based on the raw data on which they were trained. The performance of such ML models decreases greatly when another dataset is used. This problem is called the problem of generalizability [2]. NOVU uses a diverse dataset from different healthcare sources including data from several healthcare sectors such as hospitals, institutes, and radiology centres and furthermore from different countries. This makes NOVU a trustful solution work as well as advertised in any place in the world.
Single task ML model
ML models in academia and Industry are mainly focusing on firstly solving a particular problem and secondly optimizing the outcome of the algorithm. A single model is trained to perform the desired task. NOVU utilizes the concept of Multitask learning (MTL) [1] that estimates models for several tasks in a joint manner.
The black-box effect
The AI black box problem refers to the fact that with most AI-based tools, we don’t know how they do or what they do [3]. NOVU documents and establish controls across the ML model life cycle, including data gathering, preparation, model selection, training, evaluation, validation, and deployment.
Lack of interoperability
Nowadays, the IT architecture of most healthcare entities is still using the centric approach. There is a lack of interoperability among those entities. Suitable infrastructure is a must before developing ML applications. Such infrastructure can be based on cloud technology and should be able to handle big data, supporting the smooth exchange of data between the PACS and EHR [1]. NOVU is a web application which offers an integrated set of services in a fast and easy-to-use.
Data protection and privacy
Careful steps need to be taken into consideration while developing ML models in terms of data protection and privacy of patient data. The processing of patient data requires accessing large datasets that should be anonymized, de-identified, and encrypted to protect the data. In addition, special consent is needed to prevent patient privacy. It is necessary that the ML model should be transparent and secure for healthcare providers to monitor the effects of AI. Furthermore, providing a valid explanation of the workings of AI to the affected patients and other relevant parties such as HIPAA and GDPR.
Conclusion
The journey for developing accurate, robust, reproducible, generalizable, multi-task, secure ML models is still in its infancy. NOVU will advance this journey with its innovative technologies and world class multidisciplinary team involved in the development process.
Author: Dr. Mohamed Kalil