Enhancing Medical Image Instance Segmentation Using Histogram Equalization and Blind Deblurring: A Preliminary Study

Medical Image Instance Segmentation (MIIS) represents an advanced stage in medical image processing that aims to classify every pixel in an image and distinguish between individual objects belonging to the same class. This process produces precisely segmented outputs, each representing an anatomical structure, organ, or pathological region within the medical image [1, 2]. The segmentation results provide critical spatial information such as boundaries, locations, sizes, and counts of objects, serving as valuable references for medical diagnosis and clinical decision-making [3, 4].

However, MIIS remains technically challenging, with several factors negatively affecting segmentation accuracy. Key issues include the generally low quality of medical images, limited availability of annotated datasets, the diverse visual patterns of bodily structures in medical imaging, and various difficulties in image preprocessing stages [5-8]. Therefore, more innovative and flexible approaches are essential to produce more accurate and reliable segmentation outcomes.

In recent decades, various techniques have been investigated for MIIS, ranging from classical image processing methods such as edge detection, thresholding, region growing, and clustering to more recent machine learning (ML) and deep learning (DL) approaches [9]. Among DL-based methods, Mask R-CNN has emerged as one of the most effective frameworks for instance segmentation. As an extension of Faster R-CNN, Mask R-CNN introduces an additional branch to generate pixel-level masks for each Region of Interest (RoI), thereby enabling precise detection and delineation of individual anatomical instances, even in cases involving overlap or indistinct boundaries [10].

Building upon the success of Convolutional Neural Network (CNN)-based methods, various models have been proposed to enhance the performance of MIIS [11]. A key advantage of CNNs is their capability to extract complex features directly from raw medical images. These models can be trained under both supervised and unsupervised learning paradigms [12]. In this study, a supervised learning approach is employed, wherein the Mask R-CNN model is trained on annotated datasets to facilitate accurate identification of anatomical structures and their spatial characteristics.

Despite recent advancements, a key limitation of CNN-based models such as Mask R-CNN is their sensitivity to image quality. These models often underperform when applied to medical images characterized by low contrast, noise, or blurring common artifacts in modalities such as ultrasound and X-ray imaging [13-15]. While several studies have sought to address these challenges, they often lack a comprehensive preprocessing pipeline or fail to systematically evaluate the impact of multiple image enhancement techniques on instance segmentation performance.

 To address this gap, this study integrates a suite of image enhancement techniques as a preprocessing step before applying Mask R-CNN. Enhancement methods such as Histogram Equalization (HE), Adaptive Histogram Equalization (AHE), Contrast Limited Adaptive Histogram Equalization (CLAHE), and Blind Deblurring (BD) are employed to improve contrast and sharpness [16-22]. By integrating various preprocessing strategies with the instance segmentation capabilities of Mask R-CNN, this study aims to address the limitations identified in previous research. The contributions of this work include a comprehensive comparative evaluation of multiple image enhancement techniques on segmentation performance, the development of a customized preprocessing pipeline specifically designed for low-quality medical images, and the application of the proposed approach to fetal heart anatomical segmentation a domain that has received limited attention in the context of MIIS using enhanced Mask R-CNN frameworks. This integrated methodology is anticipated to improve the reliability and accuracy of instance segmentation in suboptimal imaging conditions, thereby supporting the advancement of MIIS applications in clinical settings.

The methodology in this study comprises several interrelated stages, beginning with data acquisition and culminating in the evaluation of segmentation results. The overall workflow is illustrated in Figure 1.

1.png

Figure 1. Workflow of fetal heart image segmentation using Mask R-CNN

3.1 Data acquisition

The initial stage in the SIIM process is data acquisition, as illustrated in Figure 1. The data used in this study were obtained from fetal echocardiography videos capturing the heart in the four-chamber view. These videos were sourced from an online platform with proper usage permission [30]. The video file was in .mp4 format, with a size of 13.7 MB, a duration of 178 seconds, and a frame rate of 30.00 fps. The entire video was then framed into two-dimensional ultrasound images, resulting in 357 images. These images include frames containing fetal heart objects some with one, two, or three fetal heart instances.

Additionally, some images did not contain fetal heart objects, or the heart objects were out of focus or blurred. Cropping was performed to ensure that the data used for the MIIS model met specific requirements for images containing multiple fetal heart objects or extraneous elements such as text. The output of the video extraction process and the resulting images are summarised in Table 2.

Table 2. Video extraction

3.2 Data preprocessing

The data preprocessing stage includes selection, cropping, and resizing of images, aimed at producing a dataset that represents explicitly the fetal heart structure in normal conditions. As a result of this process, a dataset consisting of 176 ultrasound images of normal fetal hearts was compiled. This dataset was divided into two main subsets: 140 images were allocated for training, while 36 images were designated for validation. Each image in this dataset represents ten primary anatomical classes of the fetal heart: Left Atrium (LA), Right Atrium (RA), Left Ventricle (LV), Right Ventricle (RV), Tricuspid Valve (TV), Pulmonary Valve (PV), Mitral Valve (MV), Aortic Valve (AV), Aorta (Ao), and Spine. Consequently, the training data contains 1,400 anatomical class labels, while the validation data includes 360 class labels. Detailed information regarding the data distribution is presented in Table 3.

The subsequent stage involves applying image enhancement techniques to improve the visibility of cardiac structures and minimise the effects of blur or noise. Several methods are utilised in this process, including HE, which enhances the global contrast of the image; AHE, which adjusts the local contrast in different regions; and CLAHE, which limits excessive contrast enhancement in overly bright or dark areas. Additionally, BD reduces blur without requiring explicit knowledge of the blur kernel. The results of these enhancement techniques are presented in Figure 2.

Table 3. Dataset MIIS

2.png

Figure 2. Fetal heart image

Image quality enhancement aims to improve visual appearance to support more accurate analysis and feature extraction. In this study, several image enhancement methods are employed, including HE, as defined by Eq. (1) [19].

$H(i)=\sum_{x=1}^M \sum_{y=1}^N \delta(f(x, y)=i)$                        (1)

AHE, as defined by Eq. (2) [20],

$f^{\prime}(x, y)=\frac{(L-1)}{|W|} \sum_{k=0}^{f(x, y)} H_{W_{x, y}}(k)$                        (2)

CLAHE, as defined by Eq. (3) [31],

$f^{\prime}(x, y)=\frac{(L-1)}{|W|} \sum_{k=0}^{f(x, y)} \min \left(H_{W_{x, y}}(k), T\right)+\frac{E}{L}$                        (3)

and BD, as defined by Eq. (4) [22].

$\hat{S}=\arg \min _s\|I-\widehat{K} * S\|^2+\lambda R(S)$                        (4)

3.3 Data annotation

In parallel with the image enhancement process, all normal fetal heart images undergo annotation using ten labels that represent the anatomical features of the fetal heart. The annotation is carried out by precisely placing polygon points on the heart regions within each image. Once completed, the annotated images are exported in JavaScript Object Notation (JSON) format for further processing. In addition to enhancing image quality, this annotation process aims to generate ground truth images manually labelled data created by experts with the competence to identify anatomical features and structures in medical images. These ground truth images are the primary reference for training and evaluating image processing models. An example of the annotation result is shown in Figure 3.

3.png

Figure 3. Ground truth of annotation results

3.4 Instance segmentation with Mask R-CNN

We employed the Mask R-CNN technique for MIIS [23]. Mask R-CNN features a core architecture that utilises a convolutional neural network (CNN) backbone, such as ResNet, to extract feature maps from the input images. These feature maps are then processed by a Region Proposal Network (RPN) to generate candidate bounding boxes that indicate potential object locations, enabling the analysis to focus on specific regions of interest.

4.png

Figure 4. Mask R-CNN architecture for deep learning-based MIIS

The primary distinction between Mask R-CNN and its predecessor, Faster R-CNN, lies in using ROI Align instead of ROI Pooling. ROI Align is designed to preserve the spatial alignment of feature maps, which is crucial for improving the accuracy of detection and segmentation. This is especially important in medical imaging, where spatial distortion caused by ROI Pooling can significantly degrade the quality of the resulting segmentation [10]. The architecture and workflow of Mask R-CNN as implemented in this study are illustrated in Figure 4.

3.5 Model evaluation

To evaluate the performance of image segmentation models, three primary metrics are commonly used: Mean Average Precision (mAP), Intersection over Union (IoU), and Dice Similarity Coefficient (DSC). Among these, mAP serves as the main evaluation metric for assessing segmentation accuracy, as it accounts for the average precision across various IoU thresholds and reflects the model's capability to detect and classify objects accurately. The mAP value is computed as the mean of the Average Precision (AP) scores across all segmented object classes, as defined in Eq. (5). [32]:

$m A P=\frac{1}{N} \sum_{i=1}^N A P_i$                        (5)

where, N is the number of classes and denotes the Average Precision for the i-th class.

Furthermore, IoU measures the extent of overlap between the predicted segmentation and the ground truth. The IoU value is calculated by comparing the area of intersection between the prediction and the ground truth with the area of their union, as defined in Eq. (6) [33]:

$I o U=\frac{|A \cap B|}{|A \cup B|}$                        (6)

where, A represents the pixels in the predicted segment and B represents the pixels in the ground truth segment. A higher IoU value indicates a more accurate segmentation performed by the model.

Next, the Dice Similarity Coefficient (DSC), or Dice Score, is used to measure the similarity between two segments, specifically between the predicted segmentation and the ground truth. This metric is particularly useful in segmentation evaluation due to its higher sensitivity to small object sizes. The DSC is defined by Eq. (7) [34]:

$D S C=\frac{2|A \cap B|}{|A|+|B|}$                        (7)

where, A and B represent the pixels of the predicted and ground truth segments, respectively. The DSC value ranges from 0 to 1, with a value of 1 indicating a perfect segmentation.

3.6 Training setup

At this stage, the environment and training parameters for the Mask R-CNN model were configured to ensure an optimal and consistent training process. This phase aimed to develop an instance segmentation model capable of accurately identifying and distinguishing anatomical structures in fetal heart images, through system configuration and training parameter adjustments tailored to the characteristics of the medical imaging data. An ablation study was conducted to determine the optimal hyperparameter settings. Key parameters such as learning rate, batch size, and number of epochs were varied to assess their impact on model accuracy, convergence, and overall performance. The results indicated that the best balance between stability and efficiency was achieved with a learning rate of 0.01, a batch size of 1 due to GPU capacity limitations, and 20 epochs with 500 steps per epoch. The detailed training parameters are presented in Table 4.

Table 4. Model training setting and environment

3.7 Model segmentation design

In this study, image segmentation models were developed using uniformly controlled experimental parameters to ensure consistency and fairness in performance evaluation. The fixed parameters applied across all experiments included 500 steps per epoch, 20 training epochs, a ResNet-50 backbone, Stochastic Gradient Descent (SGD) as the optimizer, a batch size of 1, and a learning rate of 0.01. These parameter values were selected based on the default configuration of Mask R-CNN, while the batch size was adjusted to accommodate the limitations of the available GPU resources.

The independent variables in this experiment were the learning momentum (LM) values and the image enhancement techniques applied to the training data. Two momentum values, 0.7 and 0.9, were investigated. These values were chosen based on prior studies demonstrating their effectiveness in accelerating convergence and stabilizing SGD training in medical imaging contexts. A lower momentum (e.g., 0.7) allows greater sensitivity to gradient updates, while a higher momentum (e.g., 0.9) offers more stability and smoother optimization, making this range suitable for comparative performance analysis.

The main image enhancement techniques explored in this study include HE, AHE, CLAHE, and BD. In addition to evaluating each enhancement method individually, this study also examined potential synergies by applying various combinations, including HE + CLAHE, AHE + CLAHE, HE + BD, AHE + BD, HE + CLAHE + BD, and AHE + CLAHE + BD. These were also compared with no enhancement (i.e., the original images).

In total, 22 different models were trained and evaluated, representing all possible combinations of the eleven enhancement configurations with two different momentum values applied during the optimization process. This comprehensive experimental design was conceived to systematically assess the effects of both preprocessing strategies and momentum variations on instance segmentation accuracy, all under consistent training conditions. The complete list of model configurations is presented in Table 5.

Table 5. Model experimentation

Evaluation based on average IoU reveals considerable performance variation across models. Model 14 recorded the highest average IoU of 0.5887, indicating superior segmentation capability across all classes. Specifically, this model also demonstrated strong performance in segmenting the RV, RA, and AO structures, critical components in fetal heart imaging. Models 6 and 18 also showed competitive performance, with average IoU scores above 0.50, suggesting stable segmentation across various anatomical structures. In contrast, models such as Model 1 and Model 11 exhibited relatively low performance, particularly in classes like LA and AV, which may be attributed to their small size or low contrast in the original images.

The measurement results with DCS support the findings from the IoU metrics, Model 14 again ranked the highest (average DCS = 0.7032), reinforcing the conclusion that this model can produce consistent segmentation with high overlap with the ground truth. Models 18 and 13 also demonstrated strong DCS performance, with average scores above 0.68. Classes such as RV and AO generally exhibited high DCS scores across most models, likely due to their clearer morphological contours and boundaries. In contrast, classes like TV and AV tended to have lower DCS scores, indicating that these structures are more challenging to segment accurately.

In the mAP evaluation, Model 14 again demonstrated the best performance with an average score of 0.3049, reinforcing its dominance across various evaluation metrics. This indicates that the model is not only capable of producing accurate segmentation predictions but also maintains consistency across different instances. Interestingly, although some models achieved relatively high IoU or DCS values, they did not necessarily maintain similarly high performance in mAP. For example, Model 6, which recorded strong IoU and DCS scores, showed a slightly lower mAP compared to models 14 and 13. This suggests that strong spatial segmentation does not always guarantee precise instance-level detection.

Among all evaluated models, Model 14 consistently demonstrated the best performance across the three main evaluation metrics: average IoU (0.5887), DCS (0.7032), and mAP (0.3049). The success of this model can be attributed to two key factors: the use of a LM value of 0.9 and the application of a combined preprocessing technique using AHE and CLAHE. The combination of AHE and CLAHE proved effective in enhancing local contrast in fetal cardiac ultrasound images, which is crucial for highlighting boundaries of small and low-contrast structures such as AV, TV, and LA. Meanwhile, the relatively high LM value helped the model achieve stability during training while maintaining good generalisation on test data. These findings indicate that selecting appropriate image preprocessing strategies and tuning hyperparameters optimally significantly impact segmentation performance.