Peek Health is dedicated to creating PeekMed web, a sophisticated tool for orthopedic pre-surgical planning that utilizes actual medical images from patients. In order to develop effective machine learning models for the processing of these medical images, it is essential to systematically gather, manage, and analyze real medical data. To accomplish this, Peek Health follows a structured set of steps:
-
- Gather images of adult patients from the clinical partners;
- Anonymize images to remove personal information and protect patient privacy;
- Curate data to control image quality;
- Performs relevant image annotation;
- Stores images to be used in the ML development process;
- Train ML models at least with 100 datasets;
- Deploy an ML model whenever the acceptance criteria are met;
- Validate PeekMed web after ML model deployment
- Transfer PeekMed web only when PeekMed web is fully validated.
Each ML model integrated into PeekMed web undergoes thorough performance testing to ensure it aligns with established accuracy and clinical performance standards. While PeekMed web is designed to deliver dependable quantitative imaging outputs, it is important to recognize that the performance of these models can be affected by potential biases in the training data. To address this, the ML models are trained and validated using diverse datasets sourced from clinics and hospitals around the globe, effectively reducing bias and enhancing the robustness of their performance.
Peek Health is actively engaged in the development, training, testing, and validation of a range of Machine Learning models intended for integration into PeekMed web, including:
-
- segmentation: its purpose is to isolate the bones from medical images to facilitate landmark positioning and consequently improve measurement reliability. This type of ML model exists for each orthopedic subspecialty.
- landmarking: its purpose is to automatically identify and position the landmarks on bones, which will be used to perform measurements and, when relevant, plan orthopedic surgeries. In the PRE-OP state, you are able to edit landmarks in accordance with your knowledge.
- classification: its purpose is to detect and prevent human errors, therefore identifying automatically the region and side or metal (prosthesis) in a medical image. This information is used to deny the planning if the uploaded image(s) is not in accordance with the medical procedure chosen by you or to inform you regarding the possibility of the metal detection affecting the segmentation or landmarks detection ML models.
The purpose of the ML models is to enhance user experience and streamline workflow rather than to provide medical advice. Consequently, it is essential for healthcare professionals to critically review and validate the automatically segmented bones and placed landmarks.
If you need to modify a bone segmentation or identify any errors in the landmarking or classification models, you can easily request support through PeekMed web. Simply select the “Request support” option and complete the form provided. Our support team will reach out to you promptly to offer tailored solutions that facilitate your case planning.
The development of the ML models involved the use of neural networks and a variety of rigorous training techniques, including hyperparameter tuning. This careful approach is crucial for ensuring that the models are both robust and generalizable across a wide range of clinical scenarios. To validate these models, we conducted tests on independent datasets, ensuring that there was no overlap between the training and testing data. This practice is fundamental for maintaining objectivity and allows for a thorough evaluation of the models' performance across diverse patient populations.
To promote transparency and reliability, we established clear performance specifications for each ML model prior to their development. You can access only those ML models that meet the quality standards required for inclusion in PeekMed web. The performance of the ML models is defined using the following key metrics.