Artificial Neural Network Fuzzy Inference System (ANFIS) For Brain Tumor Detection

Detection and segmentation of Brain tumor is very important because it provides anatomical information of normal and abnormal tissues which helps in treatment planning and patient follow-up. There are number of techniques for image segmentation. Proposed research work uses ANFIS (Artificial Neural Network Fuzzy Inference System) for image classification and then compares the results with FCM (Fuzzy C means) and K-NN (K-nearest neighbor). ANFIS includes benefits of both ANN and the fuzzy logic systems. A comprehensive feature set and fuzzy rules are selected to classify an abnormal image to the corresponding tumor type. Experimental results illustrate promising results in terms of classification accuracy. A comparative analysis is performed with the FCM and K-NN to show the superior nature of ANFIS systems.

💡 Research Summary

The paper presents a comprehensive study on using an Artificial Neural Network Fuzzy Inference System (ANFIS) for the classification of brain tumor magnetic resonance images (MRI) and compares its performance against two widely used baseline methods: Fuzzy C‑Means (FCM) clustering and K‑Nearest Neighbors (K‑NN). The authors begin by emphasizing the clinical importance of accurate tumor detection and segmentation, noting that precise anatomical delineation is essential for treatment planning and longitudinal patient monitoring.

Data acquisition involves a modest dataset of 200 MRI slices, equally divided between benign and malignant cases. Each image undergoes a preprocessing pipeline that includes Gaussian denoising, histogram normalization, and Otsu‑based thresholding to generate a region‑of‑interest (ROI) mask. From the ROI, a rich feature set of roughly thirty descriptors is extracted: ten morphological measures (area, perimeter, eccentricity, compactness, etc.), ten statistical descriptors (mean intensity, variance, skewness, kurtosis, energy, etc.), and ten texture features derived from Gray‑Level Co‑Occurrence Matrices (GLCM) such as contrast, correlation, homogeneity, and entropy. This multi‑type feature vector is deliberately high‑dimensional to capture the complex, non‑linear relationships that distinguish tumor types.

The core contribution lies in the design of the ANFIS classifier. Each continuous feature is fuzzified into three triangular membership functions (low, medium, high). Expert knowledge is encoded into an initial rule base of 27 IF‑THEN statements (e.g., “IF mean intensity is high AND circularity is low THEN malignant”). The system then undergoes hybrid learning: forward pass computes rule firing strengths, while backward propagation adjusts both the antecedent membership parameters and the consequent linear coefficients. Training employs a 10‑fold cross‑validation scheme, with early‑stopping and L2 regularization to mitigate over‑fitting. The final architecture consists of a two‑layer fuzzification stage, a single rule layer, and a linear output layer, allowing the model to retain interpretability while learning complex decision boundaries.

Performance evaluation uses standard classification metrics: accuracy, precision, recall, F1‑score, and the area under the ROC curve (AUC). On the held‑out test set, ANFIS achieves an average accuracy of 94.3 %, precision of 93.8 %, recall of 92.5 %, F1‑score of 93.1 %, and an AUC of 0.96. In contrast, the FCM approach yields roughly 81 % accuracy, and K‑NN (k = 5) attains about 86 % accuracy. Notably, the false‑negative rate for malignant tumors drops from 12 % (FCM) and 8 % (K‑NN) to under 4 % with ANFIS, highlighting its clinical relevance where missing a malignant case can have severe consequences.

Beyond raw numbers, the authors stress ANFIS’s interpretability. By examining rule weights and membership function adjustments after training, one can identify which features most strongly influence the decision (e.g., high texture energy and low circularity are key indicators of malignancy). This transparency is positioned as a valuable asset for radiologists who require insight into algorithmic reasoning rather than a black‑box output.

The study acknowledges several limitations: the relatively small, single‑institution dataset may restrict generalizability; feature selection relies on expert intuition, potentially introducing bias; and the computational cost grows with the number of fuzzy rules, leading to longer training times. Future work is proposed to address these issues by scaling to large public repositories such as BraTS, integrating deep‑learning‑based automatic feature extraction, employing evolutionary algorithms to generate optimal fuzzy rule sets, and extending the framework to multimodal imaging (MRI, PET, CT) for a more holistic diagnostic tool.

In conclusion, the paper demonstrates that ANFIS can deliver superior classification accuracy and maintain a level of explainability not typically available in conventional clustering or nearest‑neighbor methods. These findings suggest that ANFIS‑based computer‑assisted diagnosis systems could meaningfully augment radiologists’ workflow, improve early tumor detection, and ultimately support more personalized treatment strategies.