Diffuse infiltrating gliomas, encompassing astrocytomas and oligodendroglial tumors, pose a significant challenge in oncology due to the difficulty in distinguishing true tumor progression (PD) from pseudoprogression (PsP), a phenomenon characterized by imaging changes mimicking progression but without clinical deterioration. This distinction is critical for appropriate treatment decisions, as unnecessary interventions can be avoided, and timely treatments initiated for genuine recurrences. Glioblastoma (GBM), a high-grade glioma, has a poor prognosis, with median survival around 15 months despite standard treatment (surgical resection, radiation therapy, and chemotherapy). Current diagnostic approaches rely heavily on MRI, but differentiating PsP from PD based on MRI alone is difficult due to the overlapping imaging characteristics. PsP, typically occurring early in the post-treatment period (often within the first 3 months), is attributed to treatment-related effects such as inflammation and edema, while radiation necrosis can appear later. The current gold standard for definitive diagnosis remains surgical biopsy, an invasive procedure with limitations in accuracy and sampling. Various MRI sequences (pre- and post-contrast T1-weighted, T2-weighted, FLAIR) have been employed, but distinguishing PsP from PD remains a significant clinical challenge. Several studies have investigated advanced MRI techniques, including 18F-FET PET, rCBV measurements using DSC MRI, diffusion and perfusion imaging, and metabolic analysis, with varying degrees of success. Computer-aided diagnosis (CAD) systems using texture analysis, radiomics, machine learning, and deep learning have also been explored, showing promise in improving diagnostic accuracy. Previous deep learning studies, such as one utilizing a CNN-LSTM model with clinical and MRI data, have demonstrated potential for discriminating PsP from PD; however, these studies may lack the comprehensive integration of multiparametric MRI data employed in the present study.

The literature reveals various attempts to differentiate pseudoprogression (PsP) from true progression (PD) in gliomas. Conventional MRI sequences, while valuable, often present overlapping features, necessitating advanced imaging techniques or computer-aided diagnosis (CAD) systems for improved discrimination. Studies have investigated the use of 18F-FET PET, relative cerebral blood volume (rCBV) measurements, diffusion and perfusion imaging, and metabolic analysis. However, these approaches often involve specialized equipment or complex analysis. CAD systems offer an alternative, integrating various imaging features and leveraging machine learning or deep learning algorithms. Texture analysis, radiomics, and machine learning models have been explored, with varying levels of success. A notable recent study investigated the use of CNN-LSTM models incorporating clinical and MRI data, achieving promising results. However, a comprehensive approach utilizing a deep learning architecture that efficiently integrates all available multiparametric MRI data was lacking.

This retrospective study analyzed 43 biopsy-proven diffuse infiltrating glioma cases (WHO grade 3 or 4) that underwent adjuvant chemoradiation therapy after gross total resection. MRI data from a time point closest to the follow-up operation confirming PsP or PD were selected. The dataset comprised five original MRI sequences (pre-contrast T1-weighted, post-contrast T1-weighted, T2-weighted FSE, FLAIR, and ADC maps) and two engineered sequences (T1post-T1pre and T2-FLAIR, created by subtracting pre-contrast from post-contrast T1-weighted and FLAIR from T2-weighted images). Two deep learning models were employed: VGG16 and CNN-LSTM. For VGG16, each MRI sequence was individually input, with the model fine-tuned using the ImageNet pre-trained weights. For CNN-LSTM, three models were trained with different sets of sequences (3, 5, and 7 modalities) used as spatial sequences. Each model consisted of convolutional layers (with batch normalization and max pooling), a flatten layer, and LSTM layers, followed by a dense layer for classification. Threefold cross-validation was performed for both models. Model performance was evaluated using accuracy, AUC, and ROC curves. All images were normalized using white-stripe normalization in R.

The VGG16 model, using single MRI sequences, exhibited limited performance, with mean accuracy ranging from 0.44 to 0.60 and mean AUC from 0.47 to 0.59. In contrast, the CNN-LSTM model demonstrated significantly improved performance. The mean accuracy ranged from 0.62 to 0.75, and the mean AUC ranged from 0.64 to 0.81, with the best performance achieved using 7 modalities (T1 pre, T1 post, T2, FLAIR, ADC, T1post-T1pre, T2-FLAIR). The increase in performance was especially noteworthy when using a set of 7 modalities which had a mean accuracy of 0.75 and a mean AUC of 0.81 (95% C.I. [0.72-0.88]). The ROC curves for both VGG16 and CNN-LSTM models visually demonstrated the improved discrimination power of the CNN-LSTM model, particularly when using all seven sequences. The boxplots of the AUC values clearly showed the CNN-LSTM model's superior performance over the VGG16 model.

The findings strongly support the hypothesis that integrating all available MRI sequences into a CNN-LSTM model improves the diagnostic accuracy for differentiating PsP from PD in diffuse infiltrating gliomas. The superior performance of the CNN-LSTM model compared to the VGG16 model underscores the advantage of leveraging the temporal information inherent in the multiparametric MRI dataset. The CNN-LSTM architecture effectively captures the correlations and patterns across various sequences, enabling better discrimination of PsP from true PD. This study contributes to the growing body of evidence supporting the application of deep learning in medical imaging for improving diagnostic accuracy and potentially facilitating more informed treatment decisions. The relatively high AUC of 0.81 under the best CNN-LSTM model demonstrates the potential clinical impact of the proposed approach.

This study demonstrates the feasibility of using a CNN-LSTM model with multiparametric MRI data as a spatial sequence input for discriminating between pseudoprogression and true tumor progression in diffuse infiltrating gliomas. The CNN-LSTM model significantly outperformed the VGG16 model, highlighting the benefits of integrating multiple MRI sequences. This approach offers a promising non-invasive method for improving diagnostic accuracy and potentially reducing the need for invasive biopsies. Future research should focus on validating these results using larger datasets and exploring potential refinements of the model architecture to enhance performance and reduce computational cost.

This study is limited by its retrospective nature and relatively small sample size, particularly for PsP cases (n=7). The computational time for training increased with the number of modalities used. The results might not be generalizable to other populations or institutions due to variations in MRI protocols and image acquisition techniques. Future studies with a larger, more diverse cohort are needed to further validate the findings and assess the generalizability of the proposed method.