One of the more prevalent and severe cancer kinds is thought to be skin cancer. The main objective is to detect the melanoma in initial stage and save millions of lives. One of the most difficult aspects of developing an effective automatic classification system is due to lack of large datasets. The data imbalance and overfitting problem degrades the accuracy. In this proposed work, this problem can be solved using a Generative Adversarial Network (GAN) by generating more training images. Traditional RNNs are concerned with overcoming memory constraints. By using a cyclic link on the hidden layer, these models attain Long short-term memory. However, RNNs suffer from the issue of the gradient disappearing, which affects learning performance. To overcome these challenges this work proposes Bidirectional Long Short-Term Memory (Bi-LSTM) deep learning framework for skin cancer detection. The dataset which is collected from the International Skin Imaging Collaboration were used in image processing. A novel metaheuristic enthused by the routine of African vultures is proposed in this proposed work. The African Vulture Optimisation Algorithm (AVOA) algorithm is designed to select optimum feature of skin image. The accuracy of the proposed method obtains 98.5%. This comprehensive framework, encompassing GAN-generated data, Bi-LSTM architecture, and AVOA-based feature optimization, contributes significantly to enhancing early melanoma detection.