ABSTRACT
OBJECTIVES: This work aims to explore the impact of multicenter data heterogeneity on deep learning brain metastases (BM) autosegmentation performance, and assess the efficacy of an incremental transfer learning technique, namely learning without forgetting (LWF), to improve model generalizability without sharing raw data. MATERIALS AND METHODS: A total of six BM datasets from University Hospital Erlangen (UKER), University Hospital Zurich (USZ), Stanford, UCSF, New York University (NYU), and BraTS Challenge 2023 were used. First, the performance of the DeepMedic network for BM autosegmentation was established for exclusive single-center training and mixed multicenter training, respectively. Subsequently privacy-preserving bilateral collaboration was evaluated, where a pretrained model is shared to another center for further training using transfer learning (TL) either with or without LWF. RESULTS: For single-center training, average F1 scores of BM detection range from 0.625 (NYU) to 0.876 (UKER) on respective single-center test data. Mixed multicenter training notably improves F1 scores at Stanford and NYU, with negligible improvement at other centers. When the UKER pretrained model is applied to USZ, LWF achieves a higher average F1 score (0.839) than naive TL (0.570) and single-center training (0.688) on combined UKER and USZ test data. Naive TL improves sensitivity and contouring accuracy, but compromises precision. Conversely, LWF demonstrates commendable sensitivity, precision and contouring accuracy. When applied to Stanford, similar performance was observed. CONCLUSION: Data heterogeneity (e.g., variations in metastases density, spatial distribution, and image spatial resolution across centers) results in varying performance in BM autosegmentation, posing challenges to model generalizability. LWF is a promising approach to peer-to-peer privacy-preserving model training.
Subject(s)
Brain Neoplasms , Deep Learning , Humans , Brain Neoplasms/secondary , Brain Neoplasms/radiotherapy , PrivacyABSTRACT
Background and purpose: Magnetic resonance imaging (MRI) scans are highly sensitive to acquisition and reconstruction parameters which affect feature stability and model generalizability in radiomic research. This work aims to investigate the effect of image pre-processing and harmonization methods on the stability of brain MRI radiomic features and the prediction performance of radiomic models in patients with brain metastases (BMs). Materials and methods: Two T1 contrast enhanced brain MRI data-sets were used in this study. The first contained 25 BMs patients with scans at two different time points and was used for features stability analysis. The effect of gray level discretization (GLD), intensity normalization (Z-score, Nyul, WhiteStripe, and in house-developed method named N-Peaks), and ComBat harmonization on features stability was investigated and features with intraclass correlation coefficient >0.8 were considered as stable. The second data-set containing 64 BMs patients was used for a classification task to investigate the informativeness of stable features and the effects of harmonization methods on radiomic model performance. Results: Applying fixed bin number (FBN) GLD, resulted in higher number of stable features compare to fixed bin size (FBS) discretization (10 ± 5.5 % higher). `Harmonization in feature domain improved the stability for non-normalized and normalized images with Z-score and WhiteStripe methods. For the classification task, keeping the stable features resulted in good performance only for normalized images with N-Peaks along with FBS discretization. Conclusions: To develop a robust MRI based radiomic model we recommend using an intensity normalization method based on a reference tissue (e.g N-Peaks) and then using FBS discretization.
ABSTRACT
Defining the exact histological features of salivary gland malignancies before treatment remains an unsolved problem that compromises the ability to tailor further therapeutic steps individually. Radiomics, a new methodology to extract quantitative information from medical images, could contribute to characterizing the individual cancer phenotype already before treatment in a fast and non-invasive way. Consequently, the standardization and implementation of radiomic analysis in the clinical routine work to predict histology of salivary gland cancer (SGC) could also provide improvements in clinical decision-making. In this study, we aimed to investigate the potential of radiomic features as imaging biomarker to distinguish between high grade and low-grade salivary gland malignancies. We have also investigated the effect of image and feature level harmonization on the performance of radiomic models. For this study, our dual center cohort consisted of 126 patients, with histologically proven SGC, who underwent curative-intent treatment in two tertiary oncology centers. We extracted and analyzed the radiomics features of 120 pre-therapeutic MRI images with gadolinium (T1 sequences), and correlated those with the definitive post-operative histology. In our study the best radiomic model achieved average AUC of 0.66 and balanced accuracy of 0.63. According to the results, there is significant difference between the performance of models based on MRI intensity normalized images + harmonized features and other models (p value < 0.05) which indicates that in case of dealing with heterogeneous dataset, applying the harmonization methods is beneficial. Among radiomic features minimum intensity from first order, and gray level-variance from texture category were frequently selected during multivariate analysis which indicate the potential of these features as being used as imaging biomarker. The present bicentric study presents for the first time the feasibility of implementing MR-based, handcrafted radiomics, based on T1 contrast-enhanced sequences and the ComBat harmonization method in an effort to predict the formal grading of salivary gland carcinoma with satisfactory performance.