Ultrasound (US) is one of the most widely used imaging modalities in diagnostic and sur- gical settings due to its affordability, safety, and non-invasive nature. However, US images are prone to speckle noise, leading to low resolution and making clinical interpretation chal- lenging. Recently, researchers have applied state-of-the-art deep learning (DL) algorithms from the field of computer vision to the clinical domain. These algorithms require extensive annotated data to achieve meaningful results. In clinical US imaging, however, there are limited available datasets because annotating US images is time-consuming and requires ex- pert radiologists. Additionally, many hospitals restrict data sharing due to patient privacy policies, further limiting the development of DL algorithms for clinical US images. To ad- dress these limitations, this thesis focuses on developing innovative DL algorithms capable of performing with small datasets. Specifically, in Chapter 2, we use simulated US images as an alternative dataset to pre-train a breast tumor segmentation model. We further explore how network design complexity affects segmentation performance with limited data. In Chapter 3, we leverage 2D planes from 3D uterus US scans to develop a segmentation model using data from only 10 cervical cancer patients. In Chapter 4, we create a compact segmentation network with just 0.82 million parameters, applying knowledge distillation to transfer knowl- edge from a well-trained teacher model with 96 million parameters. This approach is ideal for portable US devices, where computational and memory-efficient models are required at the bedside. In Chapter 5, we introduce a novel approach to breast lesion classification by incorporating background as an additional class, improving the detection of invasive ductal carcinomas. In Chapter 6, we develop a framework for detecting quadriceps muscle thickness in US images, an important biomarker for frailty assessment. This framework also provides activation maps, highlighting the model’s focus on either the muscle body or bone surface. The availability of well-annotated datasets for DL model development has been a significant challenge in this thesis. To address this gap, in our final chapter, Chapter 7, we present a publicly available, expert-annotated dataset of intra-operative US images for brain tumor resection—the first of its kind, verified by two expert surgeons. Finally, in Chapter 8, we summarize our findings with concluding remarks and potential future works.