Neural Network Analysis of Non-Neoplastic Liver Histology

Full Text

To identify relevant publications, the search engines Google Scholar and PubMed were utilized. A similar search query was formulated for each platform:

["Artificial Intelligence" OR "Machine Learning" OR "Deep Learning"] AND [Liver] AND [Histology OR Microscopy OR Histopathology OR Histological] AND [Inflammatory OR Degenerative OR Hepatitis OR Cirrhosis OR Fibrosis],

and

["Artificial Intelligence"[Title/Abstract] OR "Machine Learning"[Title/Abstract] OR "Deep Learning"[Title/Abstract]] AND ["Liver"[Title/Abstract]] AND ["Histology"[Title/Abstract] OR "Microscopy"[Title/Abstract] OR "Histopathology"[Title/Abstract] OR "Histological"[Title/Abstract]] AND ["Inflammatory"[Title/Abstract] OR "Degenerative"[Title/Abstract] OR "Hepatitis"[Title/Abstract] OR "Cirrhosis"[Title/Abstract] OR "Fibrosis"[Title/Abstract]].

The search was limited to publications dated between January 1, 2020, and January 1, 2025. The initial screening yielded 160 publications. After removing duplicates and reviewing abstracts, the list was narrowed down to 56 sources.

In the next stage, the following were excluded from the sample: review articles, studies in which AI methods were used for research purposes but were not themselves the subject of development or detailed analysis, and studies that did not involve convolutional neural networks. As a result, the final set included 25 publications.

Review articles were used to verify the completeness of the source set and to identify publicly available datasets.

2.Features for Analysis

The histoarchitecture of the liver is characterized by the predominance of a single cell type – hepatocytes. A single histological slide may contain hundreds of thousands of these cells, which, on the one hand, simplifies classification tasks, but on the other hand, presents significant challenges related to the quantitative analysis of large volumes of uniform yet morphologically variable objects.

Tissue alterations in the liver generally exhibit a gradient-like distribution, reflecting the anatomical organization of the hepatic lobule. For example, in toxic injuries, the most pronounced changes are typically observed in hepatocytes adjacent to the central vein, whereas degenerative processes caused by hypoxia or nutrient deficiency primarily affect peripheral zones closer to the portal tracts.

Among the most diagnostically significant features traditionally assessed in non-neoplastic liver pathology are steatosis and balloning degeneration, hepatocellular necrosis, inflammatory infiltration (including lymphocytes, neutrophils, and eosinophils), and fibrosis – the replacement of parenchyma with connective tissue, leading to loss of organ function. These features may coexist within a single specimen, forming complex morphological patterns.

Standardized semi-quantitative scoring systems such as NAS [6], Metavir [7], and Ishak [8] are commonly used in both clinical diagnostics and research settings. These systems provide structured labels that can serve as a valuable basis for algorithmic analysis.

Thus, the presence of a dominant cell type with a limited yet clinically significant range of morphological alterations makes liver histology a promising domain for computer vision algorithms. The primary technical challenges include the gradient distribution of changes, their co-occurrence, and heterogeneity across the microscopic slide, necessitating the development of algorithms capable of capturing spatial dependencies and local contextual information.

3.Datasets

Training neural networks for the analysis of histological images requires large volumes of high-quality annotated data. Existing microscopic datasets are considerably smaller than foundational benchmark datasets that contain millions of images across dozens or hundreds of classes [9, 10]. Medical datasets, in particular, are markedly limited in scale [5, 11]. Nevertheless, such limited volumes may still suffice for model fine-tuning.

For instance, the publicly available TCGA dataset contains over 11,000 records, of which approximately 1,500 pertain to liver tumors; however, only 377 histological images are included [12]. Another dataset, PAIP 2019 [13], comprises 100 images annotated according to two classes.

For non-neoplastic liver pathology, data availability is even more restricted. Heinemann F. et al. published two datasets containing 467 and 258 liver histological images, respectively. These images were annotated either through whole-slide evaluation by a pathologist [14] or through tile-level semi-quantitative scoring using the NAS scale [15], which includes assessments of fibrosis, steatosis, ballooning, and inflammatory changes.

A larger dataset, Open TG-GATEs [16], includes over 25,000 digital liver slides, but its applicability is limited due to repeated observations typical of toxicological studies. Roy M. et al. made publicly available a dataset used in their study [17], consisting of 36 whole-slide images with individual hepatocytes exhibiting steatosis segmented.

Thus, the currently available open-access datasets are scarce, which constrains research on the application of neural network algorithms to microscopic images of non-neoplastic liver disease. Promising strategies to address this limitation include data augmentation, synthetic image generation, and the integration of multiple data sources.

4.Contextual Information in Models

Modern medical research relies on a wide range of diagnostic methods, each providing information at different levels. Consequently, there is growing interest in developing models trained on combined data that integrate heterogeneous sources of information [11]. This approach is particularly important in the analysis of non-neoplastic liver pathology, as it enables consideration of contextual parameters (e.g., biochemical, clinical, and anamnestic) associated with morphological tissue alterations [18, 19].

Despite the evident advantages of using multimodal data, studies applying such approaches to non-neoplastic liver diseases remain scarce. For instance, the Open TG-GATEs database [16], which includes histological images, molecular profiles, and clinical parameters, could theoretically support the development of comprehensive models. However, it has been utilized in only one of the studies identified [20].

At the same time, several studies have demonstrated the potential of combining histological data with other diagnostic modalities.

Bosch et al. [18] showed that integrating histological data with the portal vein pressure gradient improved the accuracy of portal hypertension assessment in patients with cirrhosis. Incorporating biochemical parameters into the model increased the AUROC from 0.76 to 0.85, confirming the enhanced predictive capacity achieved through multimodal data integration.

Jana et al. [19] proposed a method for assessing fibrosis and non-alcoholic fatty liver disease activity that combined histological data with CT imaging. The results showed an improvement in AUC by 2–14% compared to models based solely on histology, and by 20–40% compared to models using only CT data. The authors also compared different feature fusion strategies (mid-fusion vs. late-fusion) and model training approaches (single-loss vs. multi-loss) in a medical segmentation task. Their findings indicated that mid-fusion combined with a single-loss function achieved higher classification accuracy by leveraging more informative local features and promoting consistent training across modalities.

These findings demonstrate that combined models can account for contextual characteristics of disease, leading to improved diagnostic accuracy and better clinical outcomes. Incorporating heterogeneous data such as imaging, functional measurements, and biochemical parameters enables models to generate more informative predictions – particularly important in cases where histological analysis alone is insufficient to determine disease stage.

Despite the clear advantages of integrating heterogeneous data, open questions remain regarding the optimal methods for fusing information from different sources and the need to standardize data annotation approaches.

Thus, the use of multimodal data in training neural network models for the diagnosis of non-neoplastic liver diseases represents a promising direction but requires further research to refine algorithms and improve integration techniques.

5.Preprocessing and Augmentation Methods for High-Resolution Medical Images

Microscopic image datasets possess several characteristics that significantly influence the selection and application of preprocessing and augmentation techniques. These include their specific storage format (Whole Slide Image, WSI), limited dataset size, and the high cost of data acquisition [4, 5, 21]. These factors necessitate a meticulous approach to medical image processing.

In modern pathology, datasets are typically composed of high-resolution histological slide digital images, enabling detailed visualization of tissue and cellular structures. However, the large size of these images presents substantial challenges for training deep neural networks due to demanding hardware requirements [5, 21, 22].

One approach to addressing the issue of large image size is to divide WSIs into smaller fragments, or "patches" [21, 22, 23]. Williams DKA et al. [24] conducted a comparative analysis of U-Net model performance when trained on downsampled WSI images versus patch-based datasets of varying sizes and degrees of overlap. The results demonstrated that neural networks trained on patches achieved performance levels comparable to those trained on full-size images, while requiring significantly less computational power. Thus, patch-based processing of WSIs represents an effective strategy for reducing computational load without compromising model accuracy.

In the reviewed literature, information on patch size was reported in 18 studies [14, 15, 17, 19, 20, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36], with square side lengths ranging from 60 pixels [27] to 1024 pixels [27, 31].

The limited volume of histopathological datasets is primarily due to the high cost of data collection and the need for annotation by teams of highly qualified specialists [5, 21]. A promising solution to this limitation is data augmentation.

Among the reviewed studies, only nine explicitly reported the augmentation methods used: 90-degree rotations [15, 17, 26, 32, 33, 34, 37, 38], horizontal and vertical flipping [15, 17], color space transformations [31], scaling [17], and brightness adjustment [17, 26, 32, 33, 34, 37, 38]. Notably, deformative augmentation methods were not applied in the analyzed works.

In summary, analysis of existing studies reveals insufficient attention to the description of preprocessing and augmentation techniques, which hampers reproducibility and limits comparative assessment of different methodological approaches.

6.Choice of Training Methods and Loss Functions

The choice of a training method depends on the availability of annotated data and the specific diagnostic task. In the reviewed studies, the dominant approach is supervised learning, in which models are trained on manually annotated data provided by experts or by using specialized staining techniques. Alternative approaches such as weakly supervised learning are less frequently encountered, while unsupervised learning was not applied in any of the analyzed publications.

In the examined literature, supervised learning was implemented through either manual annotation by specialists [14, 25, 26, 29, 30, 31, 32, 33, 34, 38, 39, 40], machine learning algorithms such as support vector machines, decision trees, or k-nearest neighbors [17, 36, 41], or with the aid of specialized staining methods, including immunohistochemistry [18, 39, 42] and histochemistry [14, 15, 18, 25, 27, 28, 30, 34, 40, 43, 44]. These staining methods enable precise structure identification, resulting in high model accuracy – up to 100% in some studies [39, 40]. However, their practical value is limited, as such tasks can be effectively performed by trained morphologists.

Weakly supervised learning was employed in six studies [15, 19, 20, 28, 42, 43], where annotations consisted of overall NAS scores for entire slides rather than detailed region-level labeling. The number of whole-slide images used in these studies ranged from 30 [19] to 1,277 [20]. Reported model accuracies varied from 86% to 94% [15, 28, 43], depending on the diagnostic category.

Unsupervised learning was not used in the reviewed publications, although it is actively applied in other domains outside of neural network–based models [45].

Thus, supervised learning remains the most common approach in training models. However, its widespread use is constrained by the labor-intensive nature of data annotation and the subjectivity of human assessment. Weakly supervised learning appears promising, as it reduces the need for detailed annotations while maintaining high accuracy with relatively limited data. Nonetheless, the number of such studies remains small, making generalization of the results difficult. Future research directions include developing methods that minimize dependence on detailed annotations and exploring the potential of fully automated, human-free learning approaches.

The loss function is a critical component of model training, providing a quantitative measure of prediction error, stabilizing the learning process, and helping to prevent overfitting. In image segmentation tasks, loss function selection follows general principles, and no medical-specific loss functions have been established. The most commonly used functions are Dice Loss, Cross-Entropy Loss, Focal Loss, Tversky Loss, and Intersection over Union (IoU, or Jaccard) Loss, which measure agreement between predicted and ground-truth masks. In medical imaging, where class imbalance is common, specialized methods such as Focal Loss or Tversky Loss are frequently employed.

In the reviewed studies, nine research groups did not report the loss functions they used [18, 20, 38, 39, 40, 41, 42, 43, 46]. Ten studies [15, 17, 19, 26, 28, 29, 32, 33, 34, 36] employed cross-entropy or its variants, and one used the MiniMax loss function in the context of generative adversarial networks (GANs) [27]. Some studies reported the use of composite loss functions: one combined Focal Loss and Tversky Loss [31], while three others [26, 33, 34] used a combination of Smooth Loss, Softmax Cross-Entropy, and Binary Cross-Entropy.

While the reviewed loss functions have shown high performance in segmentation tasks, the predominance of cross-entropy as the default choice appears insufficiently justified. Several publications [47] suggest that specialized loss functions may yield superior results by better accounting for class imbalance. Functions such as Boundary Loss [48] and Hausdorff Loss [49], though potentially effective for precise object boundary delineation, remain rarely used and warrant further investigation. The implementation and rigorous evaluation of such specialized loss functions represent a promising direction for improving model accuracy, particularly in complex medical image segmentation tasks.

7.Neural Network Architectures

The choice of neural network architecture is a critical aspect in the development of algorithms for microscopic image analysis. Both general-purpose architectures – such as UNet, YOLO, and R-CNN [50] – and domain-specific models designed for medical imaging, such as Virchow and PLIP [11], are used in histological image segmentation and classification tasks. In the context of non-neoplastic liver pathology, architectural choices appear to be influenced by established practices and the expertise of individual research groups.

In five studies, the model architecture was not disclosed, as commercial neural network analysis platforms were used – either specialized for microscopic imaging [39, 40, 42, 43] or general-purpose services [20].

Among the reviewed studies, R-CNN [26, 32, 33, 34, 38] and U-Net [17, 31] were the most frequently used architectures for segmentation. In all six studies employing R-CNN, the authorship was nearly identical. For classification tasks, the most commonly used architectures included ResNet [18, 19, 25, 28, 29], Inception [14, 15, 18, 25], and VGG [29, 31].

Two studies [17, 34] conducted comparative evaluations of different model architectures. Hwang et al. [34] evaluated publicly available architectures including Mask R-CNN, SSD, and DeepLab, reporting superior overall performance with Mask R-CNN. Roy et al. [17] also compared standard models such as FCN and DeepLab with their own complex architectures, mostly built on U-Net variants. Through six different configurations, they identified a model that outperformed both standard and author-developed alternatives.

In addition to architectural comparisons, Baek et al. [33] assessed models trained separately on individual morphological features against a composite model trained on six features. While some single-feature models showed higher accuracy in specific tasks, the combined model achieved better overall performance. The authors concluded that multifactorial models are more suitable for clinical practice, whereas single-feature models may be valuable in research aimed at understanding pathogenesis and validating histological criteria.

Some studies also showed that traditional algorithms may outperform neural networks in diagnosing fibrosis, with logistic regression, SVM, and random forests [35], as well as decision trees and discriminant analysis [41], demonstrating comparable or superior results with lower computational demands.

Overall, the analysis reveals that the verification of non-neoplastic liver lesions predominantly relies on general-purpose convolutional neural networks pretrained on large datasets and subsequently fine-tuned on domain-specific data. The lack of specialized models in this area may reflect a prevailing trend of adapting existing architectures rather than developing novel ones. Only a few studies have compared the performance of different architectures, making it difficult to draw definitive conclusions. Moreover, in certain tasks such as fibrosis diagnosis, traditional algorithms may be preferable due to their efficiency and comparable accuracy.

No studies were found to employ specialized architectures designed explicitly for microscopic image analysis as a baseline. As a result, it is not currently possible to evaluate the effectiveness of such approaches in the context of non-neoplastic liver pathology. Future research could focus on the development and adaptation of specialized architectures that take into account the unique characteristics of liver histology slides.

Conclusion

This review highlights the substantial potential of neural network models for the analysis of microscopic images in non-neoplastic liver diseases. At the same time, it identifies significant limitations that hinder the widespread adoption of such approaches in both research and clinical practice.

First, the creation of large-scale, standardized, and high-quality annotated datasets remains a pressing challenge. Publicly available datasets on non-neoplastic liver pathology are scarce and limited in both volume and annotation depth. This constrains the development of models with high predictive accuracy and underscores the importance of directions such as synthetic data generation, augmentation, and the integration of heterogeneous data sources.

Second, most studies rely on general-purpose convolutional neural network architectures pretrained on unrelated tasks. The lack of specialized solutions tailored to the morphological features of liver tissue and semi-quantitative diagnostic tasks limits the effectiveness of such models.

Third, supervised learning remains the predominant approach; however, its implementation demands significant annotation resources. Weakly supervised learning appears to be a promising strategy for reducing dependence on fully labeled data, but relevant studies are still limited in number and scope.

Finally, the integration of contextual information – including clinical, biochemical, and radiological parameters – can significantly improve model accuracy. Nevertheless, multimodal approaches in the study of non-neoplastic liver diseases remain infrequent to date.

About the authors

Tatiana Novikova

LLC Laboratoires De Genie, Moscow, Russia

Author for correspondence.
Email: tn.path1910@yandex.ru
ORCID iD: 0000-0002-1686-5629

Врач-патологоанатом

Russian Federation

Ashot Arsenovich Melikbekyan

Email: melikbekyan.ashot@yandex.ru

Artyom M. Borbat

MVZ Pathologie Spandau

Email: aborbat@yandex.ru
ORCID iD: 0000-0002-9699-8375
SPIN-code: 8948-9169
Germany, Berlin

References

Supplementary files