Deep Learning Aids Stroke Diagnosis with CT Scans

Acute ischemic stroke (AIS) happens when blood flow to the brain is blocked. This can lead to sudden neurological symptoms that require immediate medical attention. One common way to assess stroke patients is through a type of scan called Computed Tomography (CT). However, interpreting these scans can take time and can depend a lot on the experience of the person reviewing them.

Deep Learning, a form of artificial intelligence, can help speed up the process of reading CT Scans by automatically identifying stroke Lesions. But most deep learning methods require images that have been carefully labeled, which can be a challenge. In this study, we developed a new deep learning method to find AIS lesions using CT scans from patients that were labeled but not detailed in their annotations.

#Data Collection

We gathered CT scan data from a large clinical trial known as the Third International Stroke Trial (IST-3), which included thousands of patients. These scans were taken soon after a stroke occurred and again a couple of days later. Experts reviewed these scans to label them for various brain conditions, including AIS lesions. The data consisted of scans from many hospitals, making it similar to what doctors see in everyday clinical settings.

#Preparing the Data

Before using the CT scans for training our deep learning model, we cleaned and prepared the data. This involved several steps, including converting files into a usable format, removing low-quality scans, and standardizing the size and brightness of images. We made sure to keep the data as clear as possible so that our learning model could get accurate results.

#Deep Learning Method

Our goal was to create a method to classify CT brain scans into two categories: those with AIS lesions and those without. If a lesion was present, we wanted to identify which side of the brain it was located on. To achieve this, we used a type of neural network called a convolutional neural network (CNN), which is well-suited for image analysis.

We divided our data into three parts: one for training the model, one for validating its performance, and one for testing it. This way, we could ensure that the model was learning effectively and could be reliably used on new data.

#Performance Evaluation

We evaluated how well our model could identify lesions compared to the assessments of human experts. The model showed an overall accuracy of 72% in detecting whether a lesion was present and which side of the brain it affected. Our algorithm performed better on follow-up scans, which is expected as lesions often become more visible over time.

The model also performed better when identifying larger lesions or multiple lesions. For example, it achieved 80% accuracy on larger lesions and even 100% accuracy when three or more lesions were present. However, the presence of chronic brain conditions, like old stroke lesions, made it harder for the model to classify correctly.

#Agreement with Expert Readings

To understand how our deep learning model compared with human experts, we evaluated its agreement with their assessments. The agreement level was somewhat lower for our model than for the experts themselves. This was partly because our model learned from data that included only CT scans, while experts often had access to additional imaging data for better context.

However, our model's performance closely matched that of human experts when evaluated solely on CT images, indicating that it could still provide valuable insights for stroke detection.

#Model Interpretability

Understanding how deep learning Models arrive at their decisions is crucial, especially in medical applications. To shed light on this, we used a technique called saliency mapping. This technique highlights the areas of an image that most influenced the model's predictions.

We found that the model was quite good at identifying clear AIS lesions. However, it was less certain when lesions were subtle or ambiguous, which is similar to how human interpreters might react in uncertain situations.

#Results and Findings

In our study, we analyzed over 5,700 CT scans from more than 2,300 patients. About half of these scans had visible stroke lesions according to expert readings. Our deep learning model's best performance showed that larger and multiple lesions were easier to detect, and follow-up scans offered better accuracy.

The model's performance varied based on the location of the lesions in the brain. For instance, it was more accurate at detecting lesions in certain areas like the middle cerebral artery region compared to areas like the brain stem or cerebellum, where lesions were much rarer.

#Challenges and Limitations

While we achieved good results, there were challenges. The overall accuracy was influenced by the fact that some types of lesions were less frequent in our dataset. We also noted that accuracy dipped for scans that had other chronic brain conditions present, complicating the model's ability to classify.

Additionally, one limitation was that some acute ischemic lesions might not be visible on CT scans, particularly early on. This means that labels based on the available data might not always be accurate.

#Conclusion

Our research demonstrates that deep learning can assist in identifying stroke lesions in CT scans, potentially improving the speed and accuracy of diagnosis. The model achieved an accuracy of 72% in detecting ischemic lesions and pinpointing their location. While it performed better on follow-up scans and larger lesions, several factors influenced its effectiveness, including the type and size of the lesions and the presence of chronic conditions.

The findings indicate that using large datasets from routinely collected scans can develop effective algorithms that truly represent the variety of cases seen in real-life medical settings. However, further studies are needed to improve the model's accuracy and reliability, especially regarding less common lesions and overlapping conditions.

By creating more effective ML systems based on everyday data, we can make strides toward better outcomes for patients with Acute Ischemic Strokes, ultimately enhancing the field of medical imaging.