Deep learning models for classification tasks have been widely adopted in medical imaging for diagnostic and prognostic purposes. Despite the widespread use, their internal mechanisms often remain opaque. In this paper, we present a novel evaluation technique which is designed to firstly, assess the performance of a trained classifier model, and secondly, extract heatmaps that reveal model attention in classification tasks. For this aim, a VGG-based model was trained to classify computed tomography (CT) chest images into those containing a lung abnormality, in this case Interstitial Lung Abnormalities (ILAs) vs those without ILA, using data from nearly 1000 subjects selected from one of Europes largest lung cancer screening studies (SUMMIT). The technique involves progressively replacing slices from a CT volume of one class (ILA) into a second CT volume from another class (no ILA). The technique is then asked to reclassify the CT after each slice addition. By analysing the changes in the predicted probability value (PPV) as slices from the opposing class are incorporated, we can observe the model’s shift in classification. This process is repeated across the six different planes of 3D volume (left-right, right-left, anterior-posterior, posterior-anterior, cranio-caudal, and caudo-cranial). This allows identification of specific regions within the volume that drive the model’s decision. The outcome is a heatmap of interest that highlights the critical features influencing the model’s classification from one class to another. This method provides deeper insights into the spatial features leveraged by the VGG models and highlights a volume of interest where a model focuses its attention to classify the images.