Most traditional medical diagnosis systems are founded on huge quantity of training data. However, on the occasion that very little volume of data is available, the traditional diagnosis systems derive defects such as larger error. Focused on the solution to this problem, a medical diagnosis system based on least square support vector machines (LS-SVM) is presented. Promoted training algorithm is introduced and applied on the lung cancer diagnosis system based on chest CT image. Diagnosis parameters acquisition is achieved with image processing methods, involving binarization, object selection and perimeter extraction technique in vision domain and single-level discrete 2-D wavelet transform technique in wavelet domain. Result of system training and recognition show that when limited quantity of training data are available, the system is capable of recognizing the situation and location of lung cancer. Further, the system displays superior ability of globalization to traditional systems.