With increasing advances in the field of medical brain imaging, we can now assess the presence of punctate white matter lesions (PWML) in the preterm infant. While some studies report a link between these lesions and adverse long-term outcomes, automatic detection of PWML through ultrasound (US) imaging could better assist doctors in diagnosis, at a lower cost than MRI. Many papers focus on MR biomedical image benchmark datasets, but few methods seem to tackle the detection of very small lesions in US images, because it is really challenging due to high class imbalance and low contrast. In this work, we propose a two-phase strategy: 1) Segmentation with a vision transformer to increase the number of detected lesions. 2) Multi-view classification of the lesions predicted in the output mask to reduce the number of false alarms and improve precision. We also compare 3 methods of preprocessing for input data. As a result, our method achieves better performances for PWML detection in US images compared to the best published models, with recall and precision reaching 82% and 60% respectively.