Deep learning techniques are increasingly utilized in neuroimaging evaluation, and 3D CNNs offer excellent performance in volumetric imaging. However, their reliance on big data is difficult due to the high costs and energy required to collect and describe medical data. Alternatively, 2D CNNs use 2D projections of 3D images, which frequently limits the volumetric context, affecting diagnostic accuracy. Techniques similar to transfer learning and knowledge distillation (KD) address these challenges by leveraging pre-trained models and transferring knowledge from complex teacher networks to simpler student models. These approaches improve performance while maintaining generalizability for resource-constrained medical imaging tasks.
In neuroimaging evaluation, 2D projection methods adapt 3D volumetric imaging to a 2D CNN, typically by choosing representative slices. Techniques similar to Shannon entropy have been used to discover diagnostically significant slices, while methods similar to 2D+e enhance information by combining slices. KD, introduced by Hinton, transfers knowledge from complex models to simpler ones. Recent developments include cross-modal KD, where multimodal data enhances monomodal learning, and relationship-based KD, which captures relationships between trials. However, when applying KD to train 2D CNNs, volumetric relationships in 3D imaging still need to be explored, despite its potential to improve neuroimaging classification with limited data.
Researchers from Dong-A University propose a 3D-to-2D KD framework to enhance the power of 2D CNNs to learn volumetric information from limited datasets. The framework features a 3D teacher network encoding volumetric knowledge, a 2D student network specializing in partial volumetric data, and a distillation loss to equalize feature embeddings between them. This method, when applied to Parkinson’s disease classification tasks using the 123I-DaTscan SPECT and 18F-AV133 PET datasets, demonstrated excellent performance, achieving an F1 rating of 98.30%. This projection-agnostic approach bridges the modal gap between 3D and 2D imaging, improving generalizability and addressing the challenges of medical imaging evaluation.
This method improves the representation of partial volumetric data by exploiting relational information, unlike previous approaches that depend on basic slice extraction or feature combos without specializing in damage evaluation. We introduce a “partial input restriction” strategy to upgrade KD from 3D to 2D. This includes projecting 3D volumetric data onto 2D input data using techniques similar to single slices, early fusion (channel-level pooling), joint fusion (aggregation of intermediate features), and dynamic images based on rank pooling. The 3D teacher network encodes volumetric knowledge using a modified ResNet18, and the 2D learner network, trained on partial projections, adapts to this knowledge through supervised learning and similarity-based feature matching.
The study evaluated various 2D projection methods combined with 3D-to-2D KD to improve performance. The methods included single slice input, adjoining slices (EF and JF configurations), and rank pooling techniques. The results showed consistent improvement for 3D to 2D KD conversion, with the JF-based FuseMe configuration achieving the perfect performance, comparable to the 3D tutor model. External validation of the F18-AV133 PET dataset revealed that the 2D post-KD student network outperformed the 3D teacher model. Ablation studies have highlighted the advantage of the effect of trait-based loss (Lfg) over logit-based loss (Llg). The framework effectively improved the understanding of volumetric features while eliminating modal gaps.
In summary, the study compared the proposed 3D to 2D KD approach with previous neuroimaging classification methods, highlighting the combination of 3D volumetric data. Unlike traditional CNN-based 2D systems that transform volumetric data into 2D slices, the proposed method trains a network of 3D teachers to distill knowledge right into a network of 2D learners. This process reduces computational requirements while leveraging volumetric insights for improved 2D modeling. This method is effective across a wide range of data modalities, as demonstrated with SPECT and PET imaging. Experimental results highlight its ability to generalize distribution tasks to non-distribution tasks, significantly improving performance even with limited datasets.
Check out Paper. All credit for this research goes to the researchers involved on this project. Also, remember to follow us further Twitter and join ours Telegram channel AND LinkedIn grup. If you want our work, you’ll love ours bulletin.. Don’t forget to join ours A subReddit price over 55k. ml.
Sana Hassan, an intern-consultant at Marktechpost and a dual-degree student at IIT Madras, is obsessed with applying technology and artificial intelligence to solve real-world challenges. With a powerful interest in solving practical problems, he brings a fresh perspective on the intersection of artificial intelligence and real-life solutions.
