Letter to the Editor: Traditional medical image interpretation and deep learning-based image analysis in predicting risk in patients with spontaneous intracerebral hemorrhage

Abstract

We read with great interest the study by Yang and Li in World Journal of Radiology entitled “Deep learning-based imaging model to predict early hematoma enlargement and hospital mortality in spontaneous intracerebral hemorrhage”. The advantages and disadvantages of computed tomography, magnetic resonance imaging, and deep-learning models in predicting hematoma expansion and mortality risk in patients with spontaneous intracerebral hemorrhage are discussed in this article. Manual image interpretation remains irreplaceable in emergency settings; however, its subjectivity and limitations limit its ability to predict hematoma expansion and mortality risk. Deep learning models significantly improve prediction accuracy through automated analysis and feature extraction, providing robust support for clinical decision making. However, the data requirements and complexity of these models hinder their widespread adoption. Future research should focus on optimizing model architectures, reducing data dependency, and strengthening multicenter validation to enhance their clinical applicability. Moreover, a hybrid evaluation system that combines the strengths of conventional and deep-learning approaches has the potential to enable more precise early intervention strategies for spontaneous intracerebral hemorrhage patients.

Key Words: Spontaneous intracerebral hemorrhage; Hematoma enlargement; Risk of death; Computed tomography; Magnetic resonance imaging; Deep learning model

Core Tip: In this article, we discuss the interpretation of medical images for predicting early hematoma enlargement in spontaneous intracerebral hemorrhage based on radiological features extracted through deep learning and traditional manual interpretation. We argue that artificial intelligence-based computer-aided diagnostic methods used to predict hematoma enlargement in spontaneous intracerebral hemorrhage on computed tomography images can help clinicians identify patients who would benefit from positive surgical intervention soon after admission.

TO THE EDITOR

Spontaneous intracerebral hemorrhage (SICH) is a common form of cerebrovascular disease characterized by high rates of mortality and disability. Hematoma expansion is a key contributor to clinical deterioration in patients, and the early prediction of hematoma expansion and mortality risk is essential for optimizing treatment strategies. As described in recent work by Yang and Li[1] in World Journal of Radiology, they developed a clinical-radiological model integrating clinical information with computed tomography (CT) features to predict the hemorrhage enlargement, and achieved the good results with a significant area under the curve (95% confidence interval: 0.828, range: 0.714-0.942), accuracy of 72.89%, sensitivity of 70.00%, and specificity of 74.52%. Meanwhile, in a clinical model showed a result of area under the curve was 0.534 (C-index = 0.534), overall accuracy of 55.67%, sensitivity 45.71%, and specificity 61.29%[1]. All data are superior to judgments based on clinicians' experience or simple clinical model.

Conventional imaging evaluation relies on clinician experience, which introduces subjective variability. In recent years, deep learning models have been increasingly used in medical image analysis, offering new possibilities for SICH prediction[1]. In this letter, we compare the advantages and disadvantages of CT/magnetic resonance imaging (MRI) and deep learning models for predicting hematoma expansion and mortality risk in patients with SICH, and we not only provide a broader conceptual comparison of human vs artificial intelligence (AI)-based interpretation in SICH, a perspective of hybrid diagnostic frameworks is also advocated to deepen the appliaction of AI in more diseases.

Conventional imaging assessments such as CT/MRI are the preferred methods for diagnosing SICH, as they enable the rapid identification of the bleeding site and extent. CT scans are fast and suitable for emergency scenarios; however, it is difficult to distinguish the intracerebral hemorrhage of different etiologies from CT scans alone. MRI provides richer tissue information; however, the examination time is longer than that of CT, making it unsuitable for acute-phase patients.

Deep learning methods can automatically extract radiomics features by analyzing non-contrast CT images to construct predictive models. For example, models based on the ResNet-50 convolutional neural network can identify signs of hematoma expansion that are difficult for the human eye to detect[2]. End-to-end deep learning models (e.g., U-Mamba) integrate automatic hematoma segmentation, data augmentation, and classification modules to achieve a fully automated prediction workflow[3]. These models can be integrated with algorithms such as logistic regression and naive Bayes to improve prediction accuracy.

ADVANTAGES OF MANUAL INTERPRETATION

The main advantage of expert-based interpretation is rapid diagnosis. CT scans are rapid, suitable for use in emergency scenarios, and can be used to immediately identify bleeding sites. Moreover, CT equipment is widely accessible and incurs relatively low examination costs, making it suitable for primary medical institutions. Finally, imaging signs (e.g., hematoma morphology) are straightforward and intuitive to interpret, facilitating quick assessment of patients.

DISADVANTAGES OF MANUAL INTERPRETATION

A major issue with manual interpretation is subjectivity. In addition, it is challenging to distinguish intracerebral hemorrhage with different etiologies, making it difficult to accurately predict hematoma expansion and mortality risk. The inability to perceive subtle features in images, such as tissue changes around the hematoma, can also lower manual interpretation accuracy.

ADVANTAGES OF DEEP LEARNING-BASED IMAGE ANALYSIS

The primary advantage of deep learning models is high accuracy. Moreover, end-to-end models enable full-process automation, reducing human intervention and improving efficiency. Finally, the decisions made by these models can be analyzed using techniques such as Grad-CAM, enhancing clinical trust. However, in this technique, the saliency maps do not provide enough information on explaining the accuracy of network, the relationship between classes and the modification of the images.

To demonstrate the superiority of advantages of deep learning-based image analysis, we should also clarify the distinguish between visual explanation and casual explanation. In fact, from a core logical perspective, the essence of visual explanation lies in describing the co-occurrence relationship between things or variables, without involving any judgment regarding the direction of influence. It merely describes this association. In contrast, the essence of causal explanation is to reveal the logic of mutual influence between things, clearly defining the specific direction of cause and effect and delineating the influence relationship. In terms of key requirements, visual explanation only needs to observe the association between things, without the need to exclude interfering factors or consider other disturbances such as data quality. Causal explanation, however, must satisfy three core requirements: Temporal precedence, plausible mechanism, and the exclusion of confounding variables.

In this section, a very important concept should be explained. Picture archiving and communication system stores massive amounts of data digitally through various interfaces while also providing auxiliary diagnostic management functions. It plays a crucial role in transmitting and organizing stored data among various imaging devices. The following key points should be emphasized in its integrated application. Storage tiering and cloudification, standardized interfaces and data integration, granular permissions and security protection.

DISADVANTAGES OF DEEP LEARNING-BASED IMAGE ANALYSIS

Deep learning models rely on large volumes of high-quality imaging data, and high costs are associated with data collection and annotation. Moreover, the models are complex; model construction and optimization require strong technical skills and professional expertise. Validation is also a challenge. In multi-center external validation, model performance may be influenced by differences in data distribution. Furthermore, other aspects, such as risk of overfitting in small datasets, spectrum bias, lack of standardized imaging acquisition protocols and class imbalance in mortality prediction datasets, were well discussed in our previous work[4].

Traditional manual interpretation is irreplaceable in emergency settings. However, its subjectivity and limitations restrict its ability to predict hematoma expansion and mortality risk. Deep-learning models substantially improve prediction accuracy through automated analysis and feature extraction, providing robust support for clinical decision making. However, the data requirements and complexities of these models remain barriers to their widespread adoption. Future research should focus on optimizing model architectures, reducing data dependency, and strengthening multicenter validation to enhance clinical applicability. For the robust application of AI, ethical and regulatory considerations on AI governance, which may include data privacy and security concerns, regulatory pathways, accountability in the event of diagnostic error and equity concerns (model generalizability across ethnic or geographic populations) were well discussed in previous work[4].

FUTURE DIRECTION

Future research should focus on optimizing model architectures, using federated learning to reduce data dependency, taking self-supervised or foundation models for medical imaging, deepening the application of multimodal transformers combining imaging and clinical metadata, and undergoing clinical trials to test AI-assisted decision-making. Also, we should recognize that the rigidity of the machine prohibits the complete replacement of human intellect, and AI is undergoing continuous evolution to reach the boundary of human intelligence. Humans should maintain a deeper and cautious understanding of the information AI provides before the real fusion commencement.

CONCLUSION

Manual CT/MRI interpretation and deep learning-based image analysis have different strengths and weaknesses in predicting hematoma expansion and mortality risk in patients with SICH. Traditional manual interpretation is suitable for rapid diagnosis, whereas deep-learning imaging analysis demonstrates superior performance in terms of accuracy, automation, and interpretability. By integrating the strengths of both approaches, a hybrid evaluation system could be established that offers the potential for more precise early intervention strategies for patients with SICH.