CL-MFTD: Enhanced Feature Representation with Contrastive Learning and Multiscale Filter for Tumor Detection

Ziyi Deng, Yin Zhang, Jia Liu, Fang Hu

Abstract: High-quality data and effective feature representation are critical for tumor image detection. However, existing tumor detection models are limited in relying on large-scale annotated datasets, ineffective feature extraction, and ignoring local features. This study presents an enhanced feature representation approach with contrastive learning and the multiscale filter, termed CL-MFTD, to achieve accurate tumor detection. This model realizes the pre-training via the contrastive learning framework, which extracts and compares the target and predicted features generated from different augmented views to obtain the optimal pre-trained model. Then, we investigate a multiscale adaptive Wiener filter that dynamically adjusts noise suppression strength based on local variance estimation within wavelet subbands, which overcomes the limitations of conventional single-scale Wiener filters by preserving edge structures while suppressing modality-specific noise in medical images. Finally, based on the optimal pre-trained model, a detection model is reconstructed with Backbone, Neck, and Head modules to extract effective features from the denoised dataset. We design the experiment of model comparisons and ablation tests on four evaluation metrics to verify the proposed CL-MFTD model’s performance. The results demonstrate that the CL-MFTD model outperforms other baselines and achieves satisfactory results on small-scale medical datasets.

External IDs:dblp:conf/iwcmc/DengZLH25