Malaria diagnosis is decided based on index malaria value which calculated from the amount of normal and infected erythrocyte on thin blood smear using microscope by a clinical pathologist. This activity is done manually and wastes a lot of time. Object detection using Convolutional Neural Network (CNN) is one of approach for solving this problem. However, the traditional object detection using CNN shows inadequate classification performance in labelling classes object. This paper is focused on the implementation of RetinaNet object detection approach to diagnose malaria. First, ResNet101 and ResNet50 used as RetinaNet backend network architecture for detecting both normal and infected erythrocytes on thin blood smear image with 1000x microscope zoom. Next, count every label of detected-object and calculate malaria-index value. Finally, after malaria-index value obtained, malaria diagnosis is defined. The algorithm performance with ResNet101 backend shows average precision (AP) 0,94, average recall 0,74, and average accuracy 0,73. Then the usage of ResNet50 backend in RetinaNet algorithm show average precision (AP) 0,90, average recall 0,78 and average accuracy 0,71.