Deep learning for image classification between primary central nervous system lymphoma and glioblastoma in corpus callosal tumors

INTRODUCTION

Glioblastoma (GBM) and primary central nervous system lymphoma (PCNSL) are common malignant brain tumors^[1,2] and the prognosis for these intracranial tumors is poor.^[3,4] Due to the differences in their radiological appearance, pre-operative magnetic resonance imaging (MRI) scans can help identify PCNSL from GBM. In detail, GBM is characterized by a solid homogenous enhancement with central hypointense necrosis, whereas PCNSL is characterized by a solid homogeneous enhancement.^[5,6] However, pre-operative dexamethasone injection for vasogenic edema therapy causes PCNSL necrosis, which is mistaken for GBM on MRI images.^[7] In addition, corpus callosal tumors are a challenge for classifying these two entities.

Machine learning (ML) has been utilized to distinguish PCNSL from GBM.^[8-10] However, ML remains a semi-autonomous method for classification that needs the feature extraction process from images before model development and prediction.^[10,11] Hence, the introduction of deep learning (DL) marks the beginning of a paradigm shift in image classification. This approach automatically incorporates the feature extraction and model development stages.^[12-15] Therefore, DL has been studied to distinguish these tumors using image classification. McAvoy et al. used the DL model to classify PCNSL and GBM. They reported the area under the receiver operating characteristic curve (AUC) of 0.94–0.95.,^[12] while Zhang et al. used DL for differentiating PCNSL, GBM, and tumefactive demyelinating lesions, with diagnostic performances between 0.95 and 1.00.^[14]

In clinical practice, the biopsy of corpus callosal tumors for tissue diagnosis is a challenge. The mortality rate from biopsy has been reported to be 1.1%, occurring from postoperative intracranial bleeding.^[5,16] Hence, the present study aimed to assess the diagnostic performances of the DL models between PCNSLs and GBMs at the corpus callosum.

MATERIALS AND METHODS

Convolutional neural network (CNN)

The workflow of the present study is demonstrated, as shown in [Figure 1]. A CNN architecture comprises three consecutive convolutional-pooling layers and a fully-connected classification layer. The activation function was a rectifier linear unit, and the batch size was 32. The model was trained for 30 epochs, stochastic gradient descent optimized with the Adam optimizer, and the initial learning rate set to 0.001.

Close

Figure 1:

Workflow of image classification in corpus colossal tumor.

Export to PPT

In the first baseline model, regularization and data augmentation were not performed, while regularization was conducted in the second, third, and fourth models. Ridge regression (L2) regularization and a drop-out of 0.25 were done for the second and third models, respectively. In the fourth model, data augmentation was employed using random vertical and horizontal flips, rotation, and zoom.

RESULTS

DL model development and validation

Each model reached the plateau stage after 30 epochs of training. The first baseline model was trained without regularization/data augmentation and had a testing accuracy of 0.69, but the testing loss conversely rose away from the training loss, as shown in [Figure 2a]. The 2^nd and 3^rd models have trained with the L2 and drop-out regularization, respectively. As a result, the testing accuracy of the 2^nd and 3^rd models was 0.72 and 0.73, respectively. However, the testing loss of the 2^nd model was reduced than the 3^rd model when epochs proceeded, as shown in [Figure 2b and c]. Finally, data augmentation was performed for the training of the 4^th model and the testing accuracy was 0.61, but the testing loss eventually converged to the training loss, as shown in [Figure 2d].

Close

Figure 2:

Training and testing accuracy and loss. (a) 1^st model, (b) 2^nd model with L2 regularization, (c) 3^rd model with drop-out regularization, and (d) 4^th model with data augmentation.

Export to PPT

As shown in [Table 2], the performance of all DL models for the diagnosis of PCNSL was evaluated using unviewed MRIs from the validation dataset. In terms of results, the ^1st model had sensitivity, specificity, PPV, NPV, accuracy, and F1 score of 0.59, 0.94, 0.91, 0.70, 0.77, and 0.72, respectively. When the ridge regression and drop-out regularization were performed, the accuracy of the 2^nd and 3^rd models rose by 0.83 and 0.84, respectively. At 0.68, the validation accuracy of the ^4th model was the lowest among the DL models. As shown in [Figure 3], the AUC of the 2^nd and 3^rd models had high levels of discrimination at 0.83 and 0.84, respectively, while the ^1st model had acceptable discrimination with an AUC of 0.77.

Table 2: Diagnostic performances for primary central nervous system lymphoma among models.

Model	Sensitivity (95%CI)	Specificity (95%CI)	PPV (95%CI)	NPV (95%CI)	Accuracy (95%CI)	F1 (95%CI)
1st model	0.59 (0.46–0.72)	0.94 (0.88–1.00)	0.91 (0.83–1.00)	0.70 (0.59–0.80)	0.77 (0.69–0.84)	0.72 (0.59–0.85)
2nd model	0.82 (0.72–0.92)	0.84 (0.74–0.93)	0.83 (0.74–0.93)	0.82 (0.73–0.92)	0.83 (0.76–0.90)	0.83 (0.73–0.93)
3rd model	0.75 (0.64–0.86)	0.92 (0.86–0.99)	0.91 (0.83–0.99)	0.79 (0.69–0.88)	0.84 (0.77–0.90)	0.82 (0.71–0.93)
4th model	0.75 (0.64–0.86)	0.61 (0.48–0.74)	0.66 (0.54–0.77)	0.71 (0.58–0.84)	0.68 (0.59–0.76)	0.70 (0.59–0.81)

CI: Confidence interval, PPV: Positive predictive value, NPV: Negative predictive value

Close

Figure 3:

The area under the receiver operating characteristic curve among models.

Export to PPT

DL-based web application for deployment

The 2^nd model showed a high level of diagnostic performance. As a result, a web application was constructed and launched. The tool is simple to use by scanning a quick response code or acquiring the uniform resource locator. New MRI images will be imported through computers or cell phones. The output is predicted as a diagnosis through the web application, as shown in [Figure 4].

Close

Figure 4:

Web application by the 2^nd model.

Export to PPT

DISCUSSION

In the present study, we demonstrated the possibility of using a DL model to distinguish PCNSLs from GBMs. Several methods for classifying these two cancers have been investigated, such as advanced MRIs, ML, and DL. Horger et al. compared apparent diffusion coefficient maps between these malignancies for differentiation and reported that sensitivity and specificity were 73–84% and 89–100%, respectively.^[18] Furthermore, the classification of PCNSLs and GBMs was predicted using a support vector machine (SVM) and multilayer perceptron of supervised ML, with a maximum accuracy of 69.2%. Nguyen et al. conducted a systematic review and meta-analysis of ML studies for image classification of PCNSL from GBM. They found that the reported AUC of the ML models in diagnosing PCNSL ranged from 0.878 to 0.979, and SVM was the most common algorithm used to perform image classification.^[19] However, these methods require the expertise of specialists for classification, which may cause excess workload, whereas the DL method includes image feature extraction and model development in the architecture. This approach is an alternative approach for use as a screening tool to assist clinicians in situations where specialists may be limited.

The baseline DL model without regularization has discrimination at an acceptable level. However, diagnostic performance could be enhanced by regularization with the same training dataset. L2 and drop-out regularizations improved the AUCs of models to high discrimination, while the performance of data augmentation was unacceptable in the validation process. Smirnov et al.^[20] demonstrated that dropout regularization could reduce the error rates using both testing and validated datasets for image classification, while Majanga and Viriri used dropout regularization for image segmentation of dental images and reported that the model with the dropout technique improved the accuracy of 91.3%; the model without dropout had an accuracy of 89.0%.^[21] From prior studies, data augmentation was performed to improve and help build a more reliable image classification problem using DL.^[22] Jacaruso. demonstrated data augmentation techniques in an electrocardiogram which improved performance from the accuracy of the baseline model from 82% to 88–90%.^[23]

The DL model with L2 regularization in the present study exhibited high sensitivity, making it a predictive model for use as a screening tool in clinical practice. Pewsner et al. proposed that a high-sensitivity tool is reliable when its prediction is negative because it rarely misdiagnoses people who have the condition.^[24] When the screening tool is combined with telemedicine and teleconsultant technologies, this strategy will provide significant benefits for patients in terms of effective treatment and management as well as reduced clinician workload. In a prior study, Wong and Tsai. used DL technologies with a telehealth strategy for screening and monitoring glaucoma patients.^[25]

Although the DL models in the present study had acceptable to high performance, there are some limitations that should be acknowledged. T1-Gd MRIs were used for model development in the present study; therefore, various sequences of MRIs would enhance the diagnostic performance in the future study. Moreover, we need additional images to enhance the model’s ability to differentiate PCNSL from GBM. According to experts, a good rule of thumb for image classification using DL is 1,000 images/class, with a large amount of training data.^[26] Multi-center studies will address the issue and provide more images of the specific tumor location for training the DL model.

CONCLUSION

DL with regularization has the potential to provide clinicians with reliable diagnostic performance in distinguishing PCNSL from GBM.