Letter to the Editor: Potential pitfalls in deep learning-based imaging for spontaneous intracerebral hemorrhage

Abstract

The study by Yang and Li in the World Journal of Radiology reports strong predictive performance for hematoma enlargement, perihematomal edema (PHE) and hospital mortality using hand crafted quantitative radiomics and deep learning models in patients with spontaneous intracerebral haemorrhage (ICH). While the authors report strong predictive performance using quantitative radiomics and deep learning features, several methodological concerns warrant discussion. By letting the deep learning models based on pretrained Convolutional Neural Network (CNN) to detect the density range of 50-400 Hounsfield units, may falsely detect age related and pathological basal ganglionic calcifications as heterogeneity within the ICH. This may lead to increased false positive rates in the prediction of hematoma enlargement. Similarly, by not restricting the density range between 25-35 Hounsfield units for pretrained CNN to detect PHE, may also add bias by falsely detecting age related micro ischaemic areas and chronic lacunar infarcts in basal ganglionic region as PHE. Further, hand crafted quantitative radiomics by trained clinical radiologists will calculate volume of ICH and PHE more accurately as compared to the CNN models using only three consecutive axial slices centered on the maximum haematomal cross-sectional area, as reported in the original study. Hence, an integrated model consisting of hand crafted quantitative radiomics by trained clinical radiologist with deep learning models based on pre trained CNN with the above-mentioned enhancements and a human touch by a trained clinical radiologist to remove the mentioned confounding factors is highly recommended for better outcomes.

Key Words: Intra-cerebral haematoma; Haematoma enlargement; Peri haematomal oedema; Convolutional nurolon networks; Basal ganglionic calcifications; Deep learning

Core Tip: This critique evaluates a study using radiomics and deep learning to predict outcomes in spontaneous intracerebral hemorrhage. While the original study shows promise, it may be limited by confounding calcifications and chronic ischemic lesions being misidentified as hematoma or edema due to broad Hounsfield units thresholds. Furthermore, the use of limited axial slices rather than full volumetric analysis may reduce accuracy. We propose an integrated approach that combines deep learning efficiency with strict Hounsfield units filtering and expert radiological oversight to minimize false positives and improve predictive precision.

TO THE EDITOR

We have carefully reviewed the retrospective cohort study by Yang and Li[1] recently published in the World Journal of Radiology, which investigated deep learning (DL) based models to predict hematoma enlargement (HE), perihematomal edema (PHE), and hospital mortality (HM) in patients with spontaneous intracerebral hemorrhage (ICH). The authors provided robust evidence that quantitative radiomics features from hematoma and PHE regions on non-contrast computed tomography showed good performance for predicting HE and HM. This artificial intelligence based computer-aided diagnosis method is effective for clinical decision-making regarding surgical interventions. While this research is a vital contribution, we believe the predictive framework could be significantly optimized by addressing certain pitfalls not fully explored in the original study.

TECHNICAL CHALLENGES IN DENSITY DETECTION

The primary concern involves the Hounsfield units (HU) thresholds utilized for training Convolutional Neural Networks (CNNs). Heterogeneity within the ICH, changes in hematoma volume, and the volume of PHE are critical computed tomography features for predicting re-bleed and expansion[2,3]. The original study utilized a density range of 50-400 HU for detecting ICH heterogeneity. However, the established density of ICH typically ranges from 25-88 HU[4]. Conversely, the incidence of age-related basal ganglionic calcifications (BGCs) in patients over age 50 is found to be in the range of 3.5% to 14.6%[5-9]. These calcifications have a density range of 50-400+ HU, usually exceeding 100 HU[9].

Modern DL models, such as the dual-stream feature-fusion attention U-Net model, often integrate texture, contextual, spatial and temporal features apart from simple density filtering[10]. Gradient-Based Joint Histogram Equalization (GBJHE) used for segmentation of ICH is a technique that aims to enhance the contrast of medical images. The initial operation of GBJHE is to calculate the magnitude of the gradient in the image. The gradient magnitude plays a critical role in identifying the boundaries of haematomas by emphasising the image’s edges and contours[10]. Hence, during initial operation of GBJHE to calculate the magnitude of the gradient in the image, the density overlap, as mentioned previously, is still likely to introduce significant bias in detecting heterogeneity within the ICH by misinterpreting BGCs as part of the ICH. It has also been reported in few studies that not all the pre-trained CNN could differentiate between BGCs and ICH in basal ganglion regions[11]. This may lead to increased false positive rates in prediction of HE by CNN leading to unnecessary surgical interventions. Hence, we would strongly propose while training the DL based CNN for the detection of the heterogeneity within the ICH, the density range be set in the range of 20 HU to 90 HU.

CONSTRAINTS IN PERIHEMATOMAL OEDEMA AND VOLUME ASSESSMENT

Similar issues arise regarding PHE measurement. The density of PHE has been found to be in the range of 24-30 HU[12]. For the measurement of PHE using pre-trained CNNs, the density range needs to be pre-decided (ideally 20-35 HU), which was not done in the original study. There is a high incidence of micro-vascular ischemic changes and chronic lacunar infarcts involving the basal ganglionic region in patients above 50 years of age. These chronic infarcts have a density range of -5 HU to 20 HU[13]. Without a predefined range, there is a possibility that DL models will yield high false-positivity rates by falsely detecting chronic micro-infarcts as PHE.

Furthermore, the CNN model in the original study utilized only three consecutive axial slices centred on the maximum hematoma area[1]. Such limited sampling may fail to capture the full volumetric extent of the ICH and PHE, potentially skewing predictions. A few studies have found that non-contrast computed tomography-derived 2D radiomics features exhibit acceptable, comparable performance to 3D features in predicting HE, rather than using only a few selected sections[14]. In this study 3-slice patches utilised by CNN models were cropped via a radiologist-delineated bounding box, which helped to achieve superior discriminative performance, this being an integrated model. Also, the combined clinical-radiological model demonstrated superior classification performance for prediction of HE, PHE and HM of the patients. It has previously also been noted in many studies that integrated and hybrid models have higher sensitivity and specificity in predicting HE, PHE and HM in patients with ICH[15].

PROPOSED ENHANCEMENTS FOR DL MODELS

The proposed enhancements for DL models in HE, PHE and HM prediction are given in Table 1.

Table 1 Proposed enhancements for deep learning models in hematoma enlargement, perihematomal edema, and hospital mortality prediction.

Dimension	Current metric[1]	Proposed enhancement	Clinical/methodological rationale
ICH heterogeneity range	50-400 HU	25-90 HU	Precise detection of ICH heterogeneity and volume[2,3]
PHE density range	Not defined	20-35 HU	Precise calculation of PHE volume[12]
Sampling volume	Three axial slices	Full volume or all consecutive slices	Accurate prediction of volume with exclusion of confounding factors via human touch

HU: Hounsfield units; ICH: Intracerebral haemorrhage; PHE: Perihematomal edema.

CONCLUSION

In summary, though Yang and Li[1] have successfully achieved high predictive performance for HE, PHE and HM using hand crafted quantitative radiomics and deep learning models in patients with spontaneous ICH, we advocate few enhancements to alleviate the confounding factors in the original study as discussed above. Presently, there is insufficient evidence to recommend for implementation of DL models independently for detection of HE and PHE[16]. Considering all these factors it seems that, at least presently, DL models based on CNN to predict HE and PHE needs human touch by trained clinical radiologists. Trained clinical radiologists must be an integral part of handcrafted radiomics as they can differentiate the BGCs from ICH as well as chronic microvascular and chronic lacunar infarcts within the basal ganglionic region from PHE with the help of their clinical acumen. Though the DL models are more likely to create high false positive rates as explained above, these can surely be used for the triaging of patients likely having HE and increased PHE. These triaged cases by DL models with the additional human touch by the clinical radiologists and the treating physicians can better predict the HM in patient with ICH.