Class/Regression Activation Maps (CAMs/RAMs; AMs) are often embedded into Convolutional Neural Networks (CNNs) for checking activated regions on input images at estimation. CNNs sometime generate unreliable AMs, such as activated regions, are inappropriate. Because AM is calculated by stacking many feature maps generated by the final convolutional layer, when there are Anomaly Feature Maps (AFMs), unreliable AMs can be generated. For example, suppose we have a CNN that evaluates the heart. In this case, the feature maps that focus on regions unrelated to the heart (e.g., shoulders and esophagus) are AFMs. Additionally, we have a hypothesis that the estimation accuracy of CNNs is increased by removing AFMs. However, methods for automatically detecting and removing AFMs have not been sufficiently studied in previous research to improve the performance of CNNs. Therefore, we propose a method named "Removal Operation of Anomaly Feature Maps (RO-AFMs)" to automatically detect and remove AFMs. When applying an RO-AFM to the Global Average Pooling (GAP) feature vectors of a CNN, dimensions of the GAP vector are reduced. Therefore, an RO-AFM is regarded as a deep-feature selection algorithm. From the results of adopting an RO-AFM to a Regression CNN (R-CNN) for estimating pulmonary artery wedge pressure, which is one of the measurement score for representing cardiac anomaly state, improved reliability of AM and estimation accuracy were verified. A comparison of RO-AFM and the existing methods, i.e., Lasso and the Feature Selection Layer (FSL), indicated that RO-AFM performed slightly better on the estimation accuracy. The computation time required for RO-AFM to evaluate all features was 1.833 s on average, confirming that RO-AFM is a lightweight process. Therefore, RO-AFM is useful for constructing a medical CNN that emphasizes explainability (e.g., CNNs for estimating the risk of a disease or a test value from chest X-ray or computed tomography images).