We hypothesized that deep weakly-supervised learning could detect acute ischemic stroke (AIS) and hemorrhagic infarction (HI) lesions using diffusion-weighted imaging. Each image slice was assigned an annotation indicating whether or not the slice contained a lesion. The proposed method was trained on an AIS dataset using 417 patients with weakly-labeled lesions and evaluated on a dataset using 319 patients with fully-labeled lesions, which detected lesions with high accuracy. The method was trained on a HI dataset using 240 patients with weakly-labeled lesions and evaluated using 65 patients with fully-labeled lesions. Lesion detection sensitivities were 87.7% (AISs) and 86.2% (HIs).