Wheat blast caused by Magnaphorthe oryzae pathotype Triticum (MoT) is a threat to food security. High-throughput phenotyping methods are needed for the detection of symptoms and accurate identification of wheat head blast (WHB). Current methods for wheat phenotyping include visual assessment for disease severity performed by human experts. However, significant intra and interrater variability often occur. Automated, sensor-based methods with higher reliability and accuracy have the potential to maximize phenotyping. The development of a new framework is essential to assess WHB intensity. The objective of this study is to develop an accurate deep learning framework to detect WHB symptoms. To this end, spike severity was visually estimated from experiments that included eight wheat genotypes with different levels of susceptibility to WHB. Treatments included MoT inoculated and non-inoculated wheat heads. 3155 RGB images were captured from plants grown under a controlled environment room in Bolivia. Different convolutional neural networks (CNNs) were evaluated for autonomous classification of infected heads from healthy ones in the captured images. Several experiments were conducted to illustrate the capabilities as well as the limitations of the proposed approach. Experimental results showed that the classification accuracy of our methods was 89.7% which was a significant step towards autonomous high-throughput WHB phenotyping under controlled environments.