登录查看更多内容

What is the best way to train an artificial intelligence model for medical imaging?

由人工智能和领英社区提供技术支持

Artificial intelligence (AI) is transforming the field of medical imaging, enabling faster, more accurate, and more personalized diagnosis and treatment. But how do you train an AI model to recognize complex patterns and anomalies in images such as X-rays, CT scans, MRI scans, or ultrasound? In this article, you will learn about some of the best practices and challenges of training an AI model for medical imaging.

此文章中的业界达人

由社区从 7 条内容中精选。了解更多

1 Data preparation

The first step in training an AI model for medical imaging is to prepare the data that the model will learn from. This involves collecting, labeling, and preprocessing the images. Collecting enough high-quality and diverse images is crucial for the model to generalize well to new cases. Labeling the images with the correct diagnosis or annotation requires expert knowledge and can be time-consuming and costly. Preprocessing the images involves applying techniques such as resizing, cropping, filtering, augmenting, or normalizing the images to make them suitable for the model input.

添加您的观点

Minhaj Nur Alam

AI in Healthcare | Assistant Professor at UNC Charlotte
举报内容
There are two ways you can go, 1. Go for data curation if at an early stage of exploration. Have a clean dataset and train your baseline model to see if ML is feasible. Make sure to have a training dataset that has a similar distribution like your testing dataset. You can use weights if the labels are very skewed. 2. Now if you’re at a later more advanced stage, it gets interesting. It might be worth seeing how a model performs on a biased dataset. Because in real life the disease data are ~10-20% compared to non-disease data. But during training it’s important to make sure the model has seen representative data from different sub populations.

已翻译

赞
Jo?o Oliveira

Engenheiro Biomédico | Engenheiro Clínico | Gest?o de Tecnologias em Saúde | Engenheiro de Seguran?a do Trabalho
举报内容
A coleta de um conjunto robusto de imagens, que seja tanto diversificado quanto de alta qualidade, é essencial para garantir que o modelo generalize eficazmente para cenários inéditos. A tarefa de rotular as imagens com diagnósticos ou anota??es corretas demanda conhecimento especializado e pode ser um processo demorado e dispendioso. Quanto ao pré-processamento das imagens, aplicam-se técnicas como redimensionamento, corte, filtragem, aumento ou normaliza??o, visando adaptar as imagens para uma entrada otimizada no modelo. Essa fase inicial, centrada na prepara??o dos dados, estabelece as bases cruciais para o sucesso do treinamento subsequente do modelo de IA em imagens médicas.

已翻译

赞

2 Model selection

The next step is to select the appropriate model architecture and algorithm for the task. There are many types of AI models for medical imaging, but one of the most common and powerful ones is convolutional neural networks (CNNs). CNNs are composed of layers of neurons that learn to extract features and patterns from the images. CNNs can be trained from scratch, using a large amount of data, or fine-tuned, using a pre-trained model on a similar task and adapting it to a specific problem. The choice of model depends on the availability of data, the complexity of the task, and the desired performance.

添加您的观点

Kunal Saurabh

Data Scientist @ Pyramid Consulting Inc| GenAI , IoT, Computer Vision,NLP, LLM,AI
举报内容
It is always better to use Transfer Learning. Reason being these models are trained on large volume of data and very good at recognizing low level filters i.e. shape, edges etc thereby saving time on training and deployment.

已翻译

赞

3 Model training

The third step is to train the model using the prepared data and the selected model. This involves feeding the images and their labels to the model and adjusting the model parameters to minimize the error between the model predictions and the true labels. This process is also called learning or optimization. The model training can be done on a single machine or on a distributed system, using multiple CPUs or GPUs. The model training can take from hours to days, depending on the size of the data, the complexity of the model, and the computing resources.

添加您的观点

4 Model evaluation

The fourth step is to evaluate the model performance and quality on unseen data. This involves testing the model on a separate set of images that were not used for training and measuring metrics such as accuracy, precision, recall, specificity, sensitivity, or F1-score. These metrics indicate how well the model can correctly classify or detect the images or their features. The model evaluation can also include assessing the model robustness, reliability, explainability, and fairness. These aspects indicate how well the model can handle variations, uncertainties, errors, biases, or ethical issues in the data or the task.

添加您的观点

Theodore Leng, MD, MS

Stanford Retina | Informatics Director | Director of Clinical and Translational Research
举报内容
While you can leave out a certain percentage (e.g. 10-15%) of your training dataset for model validation, ideally, the validation set will be more impactful if it comes from a separate source. Additionally, the goal would be to have your validation set contain as many of these variables as possible to maximize the effectiveness of your validation exercise: different geographic locations, different demographic/ethnic composition, different imaging device manufacturers, different imaging protocols, different pupil dilation preparation parameters, different disease severity distributions, different image quality. The more differences between your training set and your validation set, the more impactful your model will be.

已翻译

赞
Theodore Leng, MD, MS

Stanford Retina | Informatics Director | Director of Clinical and Translational Research
举报内容
One additional thing: use more than one external validation dataset for even more impact and generalizability of your model.

已翻译

赞

5 Model deployment

The final step is to deploy the model into a real-world setting, where it can be used by clinicians, patients, or researchers. This involves integrating the model with the existing systems and workflows, such as electronic health records, medical devices, or cloud platforms. The model deployment also involves monitoring and maintaining the model performance and quality over time, as well as updating the model with new data or feedback. The model deployment can face challenges such as scalability, security, privacy, or regulatory compliance.

添加您的观点

6 Model improvement

The last step is not really a final one, but rather a continuous one. It involves improving the model based on the results of the evaluation and deployment, as well as the latest advances in the field. This involves refining the data, the model, or the training process, using techniques such as data cleaning, data augmentation, model optimization, model regularization, or model ensembling. It also involves exploring new models, algorithms, or applications, using techniques such as transfer learning, multi-task learning, or generative adversarial networks. The goal is to make the model more accurate, efficient, and useful for medical imaging.

添加您的观点

Kunal Saurabh

Data Scientist @ Pyramid Consulting Inc| GenAI , IoT, Computer Vision,NLP, LLM,AI
举报内容
GAN's can be used for model improvement. They can produce not before seen synthetic images from cluster of given images thereby improving model's prediction accuracy.

已翻译

赞
Matthew Cobbinah

PhD Computer Engineering, Candidate | AI Researcher | Deep Learning | GANs | Computer Vision | Medical Imaging | Software Engineering
举报内容
There are myriad strides made in AI for health, specifically in radiology. GANs happen to be one of the promising deep learning techniques that has gained popularity since its introduction in 2014. However, the GAN architecture is fraught with challenges like mode collapse, which affects the diversity of generated samples. Hence, though strides are made, there are still relevant improvements that these techniques can receive to make them more robust in their application to intelligent downstream radiology tasks.

已翻译

赞

7 Here’s what else to consider

This is a space to share examples, stories, or insights that don’t fit into any of the previous sections. What else would you like to add?

添加您的观点

Bioengineering

+ 关注

给文章评分

我们借助人工智能创建了此文章。您认为这篇文章怎么样？

很棒不太好

举报此文章

查看全部

What is the best way to train an artificial intelligence model for medical imaging?

1

2

3

4

5

6

7

1 Data preparation

2 Model selection

3 Model training

4 Model evaluation

5 Model deployment

6 Model improvement

7 Here’s what else to consider

Bioengineering

给文章评分

感谢您的反馈

更多Bioengineering相关文章

更多相关阅读内容