DFF-U-NET: Deep features fusion u-net for bone marrow cells classification using advanced feature selection techniques

U-Net is a widely used tool for precise segmentation in medical image analysis; however, it suffers from high memory usage and feature loss. This study improves U-Net to enhance the stability and efficacy of features for classifying bone marrow cells. First, the standard structure was modified by integrating additional concatenation layers between the encoder and decoder parts to create the Deep Features Fusion U-Net (DFF-U-Net). Dropout regularization and tuned hyperparameters were employed. This boosted the network’s ability to capture higher features from different layers and levels, improving classification accuracy. Finally, a classification head was added with Leaky-ReLU with an alpha value of 0.1 to prevent vanishing gradients. Secondly, three distinct Feature Selection (FS) methods, Univariate Feature Selection (UFS), Recursive Feature Elimination with Cross-Validation (RFECV), and Select from Model (SFM), were used to select the most informative features for classification, reducing the memory requirements and accelerating computation. Further, the dataset is balanced using random oversampling. We have compared standard U-Net and DFF-U-Net with other robust EfficientNetV2S classifiers. The results showed that the DFF-U-Net classification models outperformed. In addition, our model achieved a higher performance compared to state-of-the-art methods. The proposed DFF-U-Net model was evaluated with higher accuracy, 98.93 %, compared with standard U-Net and EfficientNetV2S, 96.70 % and 94.25 %, respectively. Furthermore, the achieved results of (UFS + DFF-U-Net) are 99.11 %, (SFM + DFF-U-Net) is 99.69 %, and (RFECV + DFF-U-Net) is 99.86 %. The U-Net model is a good choice for overcoming problems in other deep learning techniques, such as limited resources, time constraints, and clinical medical applications. Our code is https://github.com/Aminaalr/Enhanced-Unet-using-Feature-Selection-Methods.