A transfer learning approach for the classification of liver cancer

1 Introduction

Liver cancer is a potentially fatal condition. However, liver cancer death rates can be reduced when it is discovered early. Many functions that are crucial for good health are handled by the liver [1]. One of these functions is the transformation of food into protein, both of which are essential for digestion. Chemicals that were inhaled and may be hazardous are removed from the body. By using vitamins, carbs, and minerals kept in the liver’s reserve, it obliterates many nutrients obtained through the digestive system and restricts the release of cholesterol [2]. The tissues of the body are built of minuscule units called cells. The usual course of events is for cells to multiply and divide to produce new cells. A new cell must be created to replace an outdated or damaged one. If the tissues of damaged or dead cells that have been expelled from the body are not eliminated, nodules and tumors may form. Malignant and benign tumors can also develop in the liver. Compared to benign tumors, malignant tumors pose a greater threat to health [3]. Early detection dramatically improves liver cancer treatment options and prognosis. When liver cancer is found early, before it has spread to other organs, the chances of a successful therapy increase. Treatments that are curative, including surgery, liver transplantation, or specialized therapy like radiofrequency ablation, are more effective in curing early-stage liver cancer. In addition, early detection makes less drastic therapeutic procedures possible, enhancing the quality of life for patients. Early-stage liver cancer treatments often have fewer side effects and need shorter recovery time than advanced-stage liver cancer treatments. With prompt identification and efficient treatment, the life expectancy for those with liver cancer can be extended. Patients have a better chance of long-term survival and chronic disease treatment with an early diagnosis, The medical images offer objective, reliable proof for liver illnesses that may be thoroughly analyzed in addition to clinical symptoms and lab test results. Among the mainstays of imaging examinations, computed tomography (CT), magnetic resonance imaging, and ultrasound are regularly employed modalities [4]. However, radiologists’ expert judgment, which has low sensitivity and probably exhibits observer bias, is used to evaluate the images. The visual examination of such a large number of medical scans is also labor and time intensive. To solve this issue, computer-aided diagnostic (CAD) systems incorporating computer vision and deep learning (DL) have been successfully created in recent years. These methods make imaging interpretation easier than manual segmentation and detection of the images [5]. The DL algorithms, as the core components of artificial intelligence, are gradually applied in CAD, negating the need for manual customization in contrast to conventional machine learning methods [6,7], which are primarily based on manually created visual descriptors and typically not optimal for specific tasks. In this article, a fine-tuning model is proposed based on DL [8] and transfer learning (TL) techniques to classify the cases from CT scan. When used to classify tumors, TL [9] is a machine learning technique where a model trained on one task is adapted for a second related activity, has various benefits:

1. Increased accuracy: TL uses information from one dataset to another, improving the model’s ability to accurately classify different cancer kinds. Pretrained models have acquired helpful features from huge datasets, which can be valuable in comprehending intricate patterns in cancer data.

2. Lessened training time and resources: creating DL models from scratch may be time and resource consuming, especially for challenging tasks like cancer classification. Due to the model starting with pre-learned features, TL considerably decreases the training time and resources necessary.

3. Managing small datasets: Because of issues with data gathering and privacy, medical datasets, notably those about cancer, frequently have a small size. Due to the fact that the model learns generic features from a bigger dataset and then customizes them for the particular cancer dataset, TL enables the model to generalize well, even with tiny datasets.

4. Feature extraction and understanding: TL models can act as feature extractors by drawing attention to crucial elements in the data that help classify cancer. This not only helps with precise forecasts but also sheds light on the characteristics that are important for classifying various tumors.

5. Adaptability to multimodal data: Imaging, genetic, and clinical records are just a few examples of the various types of datasets that are frequently used in cancer research. TL models are flexible in handling multimodal datasets because they may be trained on one type of data and then quickly expanded or modified to accommodate and classify other types of data.

6. Continuous learning: As new data become available, TL models can be continuously updated and improved. This flexibility allows the model to include the most recent developments in cancer research without necessitating a total retraining from scratch, ensuring that it remains accurate and relevant.

7. Results that can be interpreted: TL models, particularly those built on convolutional neural network (CNN) architectures, can reveal which areas or aspects of images or data are crucial for classification. In medical applications, where comprehension of the assumptions underlying forecasts is key, this interpretability is essential [10,11].

In conclusion, the efficiency and effectiveness of cancer classification models are considerably increased using TL technology, making them extremely useful for cancer research and diagnosis. Therefore, the main goal of this article is to create an intelligent system that will help physicians identify the type of liver cancer. Below is a summary of the most important contributions of this study:

1. To use a new dataset from the Radiology Institute in Iraq’s Baghdad Medical City.

2. To employ three TL learning models (VGG16, ResNet50, and MobileNetV2), along with modifications to structural models, to enhance results and accuracy.

The remaining sections of the paper are organized as follows. Section 2 presents the related works, while Section 3 presents the proposed method, results and discussions are presented in Section 4, and finally, the conclusions, as well as a few suggestions for future research, are presented in Section 5.

2 Related works

This section reviews the literatures that used DL, to detect and classify liver cancer. Since DL has semi-pioneer and requires a lot of data, it can be argued that it has become a semi-pioneer in many scientific, security, and other domains, including healthcare. These are some researches pertaining to this topic in Table 1.

3 Materials and methods

This section explains the general structure of the liver cancer classification system used to enhance the performance of DL classifiers for the classification of liver cancer, as illustrated in Figure 1.

Figure 1

The general structure of the liver cancer categorization system.

3.1 Data description

The dataset for the article was gathered from a CT scan from the Radiology Institute in Baghdad Medical City, Iraq, with an average resolution of 500 × 500 pixels; the collection consists of 700 PNG images. The images are divided into three groups: normal (231 images), benign (364 images), and malignant (105 images). The types of benign are Hydatid cysts, hemangioma, and simple cysts. Types of cysts are CE1, CE3a, CE4, and CE5. Malignant liver lesions often have an ill-defined margin with the surrounding hepatic parenchyma, whereas benign masses tend to have a well-defined parenchymal interface. There are two types of benign tumors in the paper: hemangioma and cyst. The first cannot be removed surgically; only it is followed up from time to time. As for the cysts, if they are in the first, second, or third stage, they are surgically removed. If they are in advanced stages, they are not removed because they are calcified. There are two stages in cancer cases: the primary stage and the secondary stage. The first can be surgically removed; however, the second has spread and cannot be removed due to metastasis. Only chemotherapy will be effective in treating it. Figure 2 shows some of the images from each class.

Figure 2

(a) Normal liver, (b) malignant, and (c) benign.

3.4 TL

When used to classify medical images, many CNN models have been successful. Several popular CNN topologies may be applied; they can be conventional ResNet-50 [19], VGG-16 [24], and MobileNetV2 [25], which will be the main focus of the experiment. We modified the architecture of the models, although some layers are kept frozen, new fully connected layers with random parameters are used in place of the layers from the pre-trained models. We followed the global average pooling (GAP) layer [20] to reduce the spatial dimensions of the output from the base model, a fully connected dense layer with 256 neurons and ReLU activation function [21], and a final dense layer with three neurons (for each class) and SoftMax activation function [22] to output the class probabilities. After defining the model architecture, the models are compiled using the compile function with categorical cross-entropy loss function, Adam optimizer [23], and evaluation metrics such as accuracy, recall, precision, and F1-score. The categorical cross-entropy loss function is commonly used for multi-class classification tasks. The Adam optimizer is a popular gradient-based optimization algorithm that is commonly used in DL. The accuracy metric measures the fraction of correctly classified samples, while the recall metric measures the fraction of relevant instances that are retrieved. The inclusion of recall as a metric is particularly useful in imbalanced datasets, where the number of instances for each class is significantly different, this process delivered more accuracy. The new fully connected layers can learn patterns from previously taught convolutional layers through the use of a rather low learning rate. Even without initial training, our models, VGG-16, ResNet-50, and MobileNetV2, were able to detect liver cancer by applying fine-tuning. In the second step, we used a sequential model using TL with the VGG-16, ResNet-50, and MobileNetV2 architecture pre-trained on the ImageNet dataset [26]. The process described above was applied by the following models.

3.4.1 VGG-16

The VGG-16 network that is pre-trained on ImageNet is used. The model we used included all the layers in the network layers, although the layers below are kept frozen, the pre-trained models’ layers are replaced with new three-layer GAP pooling layer and two fully connected dense layers. Our optimizer was Adam, as depicted in Figure 3.

Figure 3

VGG-16 architecture.

3.4.2 ResNet-50

The ResNet-50 network pre-trained was also employed. Every layer in the network layers was included in the model we utilized. Although the layers below are kept frozen, new three-layer GAP pooling layers and two fully connected dense layers are used in place of the pre-trained models’ layers. Our optimizer was Adam as shown in Figure 4.

Figure 4

ResNet-50 architecture.

3.4.3 MobileNetV2

The pre-trained MobileNetV2 network was also used. The model we used comprised every layer in the network layers. The pre-trained model layers are replaced with new three-layer GAP pooling layers and two fully connected dense layers while the levels below are maintained frozen with the optimizer Adam as shown in Figure 5.

Figure 5

MobileNetv2 architecture.

4 Results and discussion

This section is divided into the following parts: results are provided in Section 4.1; results are compared to related works in Section 4.2; and research limitations are presented in Section 4.3.

4.1 Results

The proposed TL model for liver cancer disease diagnosis is assessed after each epoch. Three pre-trained models were utilized to identify and classify the distinctive features in each liver image. The obtained results for our dataset show a final diagnosis accuracy of 99.10% for the VGG-16 and 99.10 for the MobileNetV2 model. By comparison with the final diagnosis, accuracy was 100% for the ResNet-50 model. Metrics are summarized in Figures 6–11.

Figure 6

Classification accuracy using the ResNet-50 model, with support for recall, precision, and F1-score.

Figure 7

Training progress for ResNet-50.

Figure 8

Classification accuracy for VGG-16 model to precision, recall, and F1-score.

Figure 9

Training progress for VGG-16.

Figure 10

Classification accuracy for the MobileNetV2 model to precision, recall, and F1-score.

Figure 11

Training progress for MobileNetV2.

We utilized the following metrics: accuracy, recall, and F1-score. The below formulas were used to calculate each measurement

(1) Accuracy = TP + TN TP + TN + FP + FN ,

(2) Precision = TP TP + FP ,

(3) Recall = TP TP + FN ,

(4) F 1 ‐ score = 2 × ( Precision × Recall ) ( Precision + Recall ) ,

where TP , TN , FP , and FN are the numbers of the true positive, true negative, false positive, and false negative results, respectively.

From Figures 7, 9, and 11, it can be seen that the loss starts with high values while learning is still low and then declines to its lowest value during the most utilized epoch. The classifier has excellent performance because of the validation test’s typical behavior, which consistently classifies the training examples. When minimal learning has occurred, the loss starts out high and then decreases to its lowest value, while the learning is increasing and reaching its highest level at the most used epoch. The classifier will perform exceptionally well because of the validation test’s normal behavior, which typically identifies the training data. Additionally, the number of epochs varies from one model to another and according to the way the data set is divided. Compared to ResNet50V2 and MobileNetV2, which each utilized 13 epochs, VGG-16 used 20.

5 Conclusions and future works

The liver is in charge of several different processes that are crucial to keeping one’s health. one of these processes, The liver may develop benign or malignant tumors, two different types of growths. Expertise in this area is necessary to detect small changes in the liver. However, such minute changes might not always be seen by the human eye. A lot of lives could be saved by medical assistance made possible by DL. This motivation led to the use a DL technique, namely the TL to provide a reliable method for identifying cases of liver cancer. In this paper, a collection of procedures is employed to identify liver cancer. This paper assessed the effectiveness of TL using ResNet-50, VGG-16, and MobileNetV2. We observed that the algorithms’ best performance was attained. After contributing to altering the models’ structure by adding necessary layers and freezing others, a high level of diagnostic precision was obtained., where the proposed models using TL (ResNet-50, VGG-16 and MobileNetV2) obtained an accuracy of 100, 99, and 99%, respectively. As it turns out, modern research techniques such as TL in DL can successfully handle diagnosing issues for physicians. Furthermore, data augmentation improved the amount of training data that was available throughout the training phase, which helped the suggested model avoid the overfitting issue. For future work, based on what we have observed in the outcomes of prior studies and our own discoveries, we suggest trying to extend the dataset size while utilizing other modern algorithms of DL, such as You Look Only One (Yolo) algorithm.